Imagine the same detective working two different cases. In the first, every piece of evidence arrives as a disconnected fragment—names without relationships, events without timelines, motives without context. In the second, the connections among people, events, and evidence are already established before the detective walks in.

Which case does the detective solve faster?

The detective solves the second case much faster than the first, but not because the detective is more intelligent. Instead, the world is more intelligible. The difference is not reasoning ability; it is the organization of knowledge that represents the world and enables action.

Now, imagine two cybersecurity AI agents built on the same frontier language model, with the objective of defending against attacks. Both share a general understanding of vulnerabilities, attackers, and defensive techniques. One, however, operates over fragmented alerts, disconnected data sources, and inconsistent identifiers. The other agent reasons over a structured representation that captures relationships among assets, vulnerabilities, threat actors, detections, and organizational context. Similar to detectives, the difference in performance stems not from greater intelligence but from a more intelligible environment or even a living enterprise model.

This is my primary argument: Agentic intelligence is about constructing a representation of the operational world in which correct decisions become the expected outcome rather than a matter of chance. As agents begin taking actions rather than merely producing answers, the quality of those actions depends mostly on the context the agents inherit and whether it’s structured, current, and backed by trustworthy evidence.

To understand how this works for machines, it helps to examine human learning. Human progress has never depended on information alone. It has depended on organizing that information—deciding which sources to trust, connecting related facts, preserving context, and passing that knowledge forward. We inherit this structure through education, language, institutions, and shared practice. Remove it, and even sophisticated reasoning becomes little more than educated guesswork.

Intelligence doesn’t emerge in isolation. It depends on the quality of the representations we think with. The quality of reasoning depends not only on our ability to draw conclusions, but also on how well the world has been made intelligible before we begin. Trustworthy agents do not emerge from better prompts. They emerge from better operational world models. Humans develop reasoning by building increasingly rich mental models of the world through observation, interaction, and experience. Modern language models follow a different path. They learn statistical representations from vast amounts of data rather than through direct interaction with the environment. As a result, they possess broad knowledge but lack direct access to the current operational state of the world in which they must act.

This distinction is particularly important in cybersecurity. Human analysts continuously integrate observations, institutional knowledge, and information about operational state into a coherent mental model of their environment. AI agents must instead be provided with that operational model. Assets, identities, vulnerabilities, dependencies, relationships, policies, and evolving evidence cannot be inferred reliably from language alone—they must be represented explicitly in the world the agent reasons over. Foundation models are increasingly capable of planning, selecting tools, interpreting observations, and adapting their behavior. Their effectiveness, therefore, depends less on reasoning in isolation than on the quality of the representations they can access and the actions they can take. Those representations encompass not only organizational knowledge, like entities, relationships, provenance, policies, governance, and evolving state, but also the tools available to observe and change that state. Again, general models generate general answers.

During the development of enterprise AI agents at Recorded Future, we encountered many instructive examples. The first lesson came immediately. Early versions of our agents treated open-source information and Recorded Future’s proprietary intelligence with roughly equal weight. The result was generic, internet-scale responses rather than the expert-level analysis we intended. Through repeated experimentation, we changed the architecture so that the agent reasoned primarily over the Recorded Future Intelligence Graph®¹—a structured representation of the cyber threat landscape built from years of analyst research, enriched relationships, and investigative decisions. The agent learned to prioritize high-confidence analytical artifacts, such as Insikt Notes² and curated graph relationships, over weaker signals derived from open sources. The language model itself had not become more intelligent. What changed was the quality of the world it was given to reason about. The result was an agent that produced more authoritative analyses and acted with substantially greater consistency and confidence.

Frontier models are becoming easier to access every month. That means model choice is becoming a weaker source of competitive advantage than many organizations assume. What’s much harder to copy is a trustworthy representation of an organization’s operational knowledge. While production-grade orchestration remains challenging, it is unlikely to provide a durable moat for agentic systems in cybersecurity as frameworks, tooling, and implementation patterns continue to converge. What cannot be commoditized is a trusted operational representation of the organization. Agents struggle wherever knowledge remains implicit—in institutional expertise, undocumented conventions, or relationships no system has captured. Their performance ultimately reflects the quality of the world they can access.

Consider a seemingly simple Priority Intelligence Requirement (PIR): “What is a threat to our organization?” The answer is rarely explicit or universally agreed upon. It emerges from understanding the external threat landscape together with the organization’s own high-value assets, priorities, and risk posture. Constructing that representation is a deliberate analytical exercise, not a by-product of deploying an AI model. Once made explicit, however, it provides the foundation on which agents can reason consistently and act effectively. The differentiating question is therefore not, “Which agent framework do we use?” but, “What operational world do our agents reason over?” Organizations that invest in making their knowledge explicit, structured, and trustworthy will consistently derive more value from AI than those that deploy increasingly sophisticated models against fragmented operational knowledge. In this case, intelligence is truly an architecture rather than computation. To build a structured world for cybersecurity agents to operate in, enterprises should align with the following principles:

Structure before reasoning. An effective AI system does more than retrieve facts. Too many systems repeatedly reconstruct knowledge from unstructured information, wasting computational resources while producing inconsistent results. Instead, intelligent and trusted agents should inherit an organized representation of the world: relationships between entities, events, organizations, and evidence should already exist within a structured living knowledge system. The agent's responsibility is to reason over this established structure, not rediscover it with every interaction. The result can be greater consistency and reliable outcomes, particularly for tasks that depend on relationships across multiple sources. Whether it also lowers total cost depends on how the representation is constructed, maintained, and reused³.

We observed the same pattern when building an autonomous threat hunting capability in our Autonomous Threat Operations (ATO)⁴. The language model already understood cybersecurity concepts—it knew what YARA rules, Sigma rules, MITRE ATT&CK techniques, and behavioral indicators were. The breakthrough did not come from teaching more cybersecurity data. It came from explicitly representing investigations: translating detection rules into behavioral hunts, preserving relationships between evidence, maintaining provenance, and carrying forward prior investigative decisions. An investigation that once required analysts to repeatedly reconstruct context became dramatically faster because the context itself had become part of the system. Before the change, each run reconstructed the relationship between the rule, the relevant behavior, and prior investigative findings. After those relationships became a durable system state, the agent could begin from the current investigative position rather than restarting from the source material. As a result, the agent could act effectively around the clock at greater speed.

Provenance is inseparable from intelligence. Human experts are trained to question sources, recognize outdated information, and detect uncertainty. AI systems cannot make these judgments unless every piece of information carries evidence of its origin, relevance, and reliability. Provenance is what transforms data into trustworthy intelligence⁵. For operational intelligence, provenance should include the source; the precise evidence used and its freshness, confidence, and known contradictions; and the relationship between that evidence and the resulting claim. Without provenance, even sophisticated reasoning is reasoning built on assumptions rather than verified knowledge.

A major pain point for cyber threat intelligence (CTI) teams is that, on any given day, thousands of articles, social media posts, blog entries, and automated alerts may mention an entity of interest to them. Most are operationally irrelevant: news commentary, recycled reporting, speculative claims, or duplicate coverage. The intelligence value lies not in the volume of mentions but in understanding their provenance. Is the information derived from a trusted government advisory, firsthand reporting from an intrusion, or direct observations from a darknet forum? Or is it simply another outlet repeating yesterday’s story? Provenance enables an AI agent to distinguish original evidence from repetition, assess freshness and credibility, and focus attention on developments that materially change the threat landscape rather than amplifying noise.

Verification must be easier than investigation. Trust matters more than raw capability. An AI system that produces impressive answers but cannot explain how it reached them creates uncertainty rather than confidence. Every recommendation should be accompanied by sufficient evidence to make verification substantially easier than reproducing the investigation itself. This includes the observations used, material assumptions, uncertainty, contradictory evidence considered, and the conditions under which the recommendation should not be followed. If it doesn't, the technology has merely shifted labor instead of reducing it.

Transparency is the foundation on which trust is built. This principle increasingly aligns with emerging thinking in frontier AI governance, where trust depends not only on model capability but on rigorous verification of the evidence, processes, and claims underlying AI systems⁶.

A CISO, good friend, operating under significant resource constraints, once asked whether an AI agent could reliably recommend isolating compromised production servers. My answer was yes—but only under one condition. A threat analyst should never have to reconstruct the entire investigation just to decide whether to trust a recommendation coming from an AI agent (it is also very costly). Instead, the AI agent should present the evidence for the underlying conclusion, like the suspicious process it observed, the connection to known malicious infrastructure, the behavioral detections that fired, the supporting threat intelligence, the confidence of each finding, and any contradictory evidence it considered, but also the reasoning that connects those observations. Evidence alone is insufficient; the same facts can often support multiple competing hypotheses.

The goal is not for analysts to reconstruct every investigation from first principles, but to quickly assess whether the agent’s chain of reasoning is sound and whether the recommended action is proportionate to the available evidence. This becomes even more important as agents transition from advisory to autonomous roles. Defensive agents offered by the industry will increasingly be expected to isolate hosts, revoke credentials, or block communications before a human can intervene. In those settings, trust cannot depend on manual verification of every decision. It must be earned through consistently reliable reasoning, calibrated confidence, transparent evidence, and well-defined operational guardrails. Verifying the recommendation should take minutes; reconstructing it from raw telemetry could take hours. Otherwise, the AI has not reduced the workload—it has merely shifted it from investigation to verification.

Efficiency is an intellectual discipline. Many assume that more reasoning necessarily produces better intelligence, yet excessive computation often reflects poor system design rather than greater capability. Efficient intelligence is not about thinking longer; it is about thinking more precisely. A well-designed agent retrieves only what is relevant, avoids unnecessary repetition, and allocates computational effort where it genuinely adds value. Efficiency is not merely an economic concern—it is a measure of intellectual discipline⁷.

In my ongoing conversations with security leaders, one concern has surfaced repeatedly: Organizations cannot afford to reproduce the inefficiencies of alert fatigue with AI agents. Replacing thousands of alerts with thousands of autonomous actions would simply automate the noise. Agents are only as effective as the intelligence, policies, and evidence they operate upon. If the underlying representations contain duplicates, false positives, stale context, or conflicting evidence, autonomous systems will amplify those weaknesses at machine speed. The goal is therefore not to maximize agent activity but to maximize decision quality. The most valuable agent consistently performs the right actions while knowing when not to act.

Reasoning must be preserved, not just data. As organizations deploy multiple intelligent agents, preserving reasoning becomes as important as preserving data. What must persist is not an unrestricted transcript of prior reasoning, but a governed record of evidence examined, hypotheses considered, assumptions made, decisions taken, unresolved questions, confidence levels, and conclusions that have subsequently become stale or invalid. Intelligence is cumulative. Human collaboration succeeds because people exchange not only conclusions but also assumptions, uncertainties, confidence levels, and unresolved questions. Intelligent agents should communicate in the same manner: Each agent should inherit the complete investigative state rather than merely a final answer. This enables cumulative reasoning—each participant building on prior understanding rather than restarting the thinking process—and creates consistency, transparency, and trust across increasingly complex systems.

One of the most common cyber cases I encounter is a ransomware investigation involving multiple analysts, shifts, and, increasingly, specialized AI agents. Whether the next participant is human or machine, they should not begin with only the latest conclusion. They need to inherit the investigative state: the evidence already examined, assumptions made, confidence assigned, competing hypotheses, unresolved questions, and conclusions that have since been revised as new evidence emerged. Without that shared state, every handoff requires reconstructing why earlier decisions were made and which lines of inquiry were already explored. With it, the investigation progresses cumulatively rather than repeatedly restarting, reducing duplication while improving consistency, auditability, and decision quality.

Intelligence must outlive the interface. Technology platforms, user interfaces, and AI models will inevitably evolve, but high-quality intelligence should remain portable, structured, and universally accessible. Organizations should not become dependent on a single application or agent. Intelligence should function as a foundational layer that supports any workflow, regardless of where or how it is consumed.

Threat intelligence organizations have spent years enriching adversary profiles, linking infrastructure, documenting investigative decisions, and building behavioral knowledge—not to mention identifying all the IT debt organizations have. If that intelligence exists only inside one vendor’s application or is encoded in prompts for a single AI assistant, every platform migration forces the organization to rebuild years of accumulated knowledge. By contrast, when intelligence is represented independently (as structured entities, relationships, evidence, and provenance), it can support any analyst, detection engine, hunting agent, or future AI model without having to be recreated. Customer relationships, business policies, operational knowledge, and historical decisions should not be lost when an organization replaces its CRM, SIEM, or AI provider. The applications may change; the intelligence they rely upon should not. Durable intelligence becomes a shared asset that every future system can inherit. Every major platform migration exposes the same problem: Organizations have preserved their applications better than their intelligence.

None of this diminishes the human role; in fact, it centers it. Rather than forcing analysts to build every investigation from scratch, intelligent systems should prepare a structured foundation upon which human expertise can operate. AI becomes a collaborator rather than a substitute. The highest value of an intelligent agent is not that it thinks independently, but that it enables humans to think more effectively.

For decades, we have treated intelligence as something people consume: reports, recommendations, indicators, documents. Agentic systems are forcing us to take a different view. Intelligence has become part of the environment in which reasoning occurs. The competitive advantage will not belong to organizations possessing the largest models. It will belong to those who build trustworthy systems where knowledge is structured, evidence is transparent, reasoning is efficient, and human judgment remains central.

Models will change. Interfaces will disappear. Workflows will evolve. What remains is the architecture of enterprise intelligence itself.

In August 2026, Insikt Group® identified 73 high-impact vulnerabilities that should be prioritized for remediation, 43 of which had a Very Critical Recorded Future Risk Score. This represents a 14% decrease from last month. 31 of these vulnerabilities were surfaced through the US Cybersecurity and Infrastructure Security Agency (CISA)’s Known Exploited Vulnerabilities (KEV) catalog, 32 were reported in open sources and validated by Insikt Group, seven were sourced through security vendor telemetry, and three were exclusively surfaced through honeypot data.

The 73 vulnerabilities in this blog affected products from 45 vendors, with Microsoft accounting for approximately 11% of the vulnerabilities. The remaining exposure spanned remote monitoring and management, virtualization, application delivery, collaboration, artificial intelligence, developer, analytics, identity, operational technology, content management, network edge, video surveillance, and endpoint technologies.

In August, Insikt Group created Nuclei templates to detect CVE-2025-62593 (Ray), CVE-2026-72898 (Metabase), and CVE-2026-9198 (IBM Langflow). Each of these vulnerabilities is featured in this blog. Additionally, Insikt Group had previously created templates to detect CVE-2026-3395 (MaxSite CMS) and CVE-2026-59800 (decolua 9Router), but their exploitation was reported in July, so they are not listed in the August 2026 Vulnerability Table. Additionally, Insikt Group created a Nuclei template to detect GitHub Issue #4255 affecting Apache Log4j, a deserialization allowlist bypass that Apache classified as a hardening gap rather than a Log4j vulnerability; as such, it was not assigned a CVE. These Nuclei templates are available to customers via the Recorded Future Intelligence Platform.

Quick reference: August 2026 vulnerability table

All 70 vulnerabilities below were actively exploited or operationally weaponized in August 2026. This table does not include the three CVEs that were primarily surfaced through honeypot data, which are available to Recorded Future Intelligence Platform customers via the CVE Monthly report. The table below also provides examples of public PoCs identified by Insikt Group. These PoCs were not tested for accuracy or efficacy. Vulnerability management teams should exercise caution and verify the validity of PoCs before testing.

Table 1: List of vulnerabilities that were actively exploited in August, 2026 based on Recorded Future data (excluding honeypot-sourced CVEs).

Key trends: August 2026

• August reporting showed two AI-assisted operations: UAT-10147 exploited Zimbra, AjaxPro, Nacos, and Telerik servers before using DeepAudit and PentestGPT after compromise, while a separate Chinese-speaking actor weaponized Hermes Agent and DeepSeek in failed attempts against Langflow and n8n.
• 34 of the 73 vulnerabilities enabled remote code execution (RCE). They affected Microsoft productivity, database, server, and endpoint software; network edge and application delivery appliances; webmail, collaboration, content management, and web server platforms; AI, analytics, developer, and CI/CD services; and operational technology, product lifecycle management, file transfer, videoconferencing, and enterprise integration software.
• We identified public proof-of-concept (PoC) exploits and scanners for 53 of the 73 vulnerabilities.
• The most common weakness classes were CWE-94 (Code Injection) and CWE-502 (Deserialization of Untrusted Data) with seven each, followed by CWE-287 (Improper Authentication) and CWE-306 (Missing Authentication for Critical Function) with six each.
• 17 vulnerabilities were at least five years old, and the oldest was approximately 16 years old.

Trend analysis: AI-assisted workflows scale exploitation and target selection

Insikt Group detailed how Chinese-speaking threat group, UAT-10147, combined conventional exploitation with agentic artificial intelligence (AI) during post-compromise operations against internet-facing Windows and Linux web servers. Cisco Talos observed the threat actor exploiting or weaponizing CVE-2019-18935 in Telerik UI for ASP.NET AJAX, CVE-2021-23758 in AjaxPro, CVE-2021-29441 and CVE-2021-29442 in Nacos, and CVE-2022-27925 in Zimbra for initial access. After compromise, UAT-10147 used CVE-2010-3904, CVE-2015-3246, CVE-2015-5287, CVE-2021-3156, CVE-2022-0847, and CVE-2022-0995 for local privilege escalation on Linux.

Figure 1: Risk Rules history on the Vulnerability Intelligence Card® for CVE-2021-23758 in Recorded Future (Source: Recorded Future)

Today, Recorded Future is announcing Automated Signature Creation, a new capability in Attack Surface Intelligence (ASI) to combat the speed of AI-generated exploits.

ASI continuously maps an organization’s external exposure, correlates newly surfaced vulnerabilities with real-world threat intelligence, and prioritizes response to enable defenders to remediate before adversaries can act.

This new function within ASI automatically creates signatures, pieces of detection logic that empowers the Recorded Future Platform to recognize a specific vulnerable or exposed condition across organization’s assets in real time.

With Automated Signature Creation now available, Recorded Future is helping to close the gap between AI-enabled threat discovery and enterprise defense.

Tackling the machine speed of exploitable vulnerabilities

It seems everything is moving quicker these days and the time to exploit a vulnerability is no different. A new generation of AI models is accelerating this challenge, demonstrating that they can automatically find zero-day vulnerabilities in major operating systems and web browsers — a skill that was previously exclusive to the most advanced government cyber units and research labs.

Back in 2020, we cited how Gartner confirmed that the time from discovery to exploitation dropped from 45 days to 15 days, between 2010 to 2020.

In our 2025 Malware and Vulnerability Trends report, we reported that weaponization occurred “within days of disclosure." Today, that window is measured in hours.

As a result, the status quo of traditional defenses and manual processes are no longer sufficient. Let’s look back at how we got here, from pre-existing detection methods to Recorded Future’s latest ASI enhancement to better defend against AI-accelerated vulnerabilities.

How we got here

In the past year, Recorded Future’s traditional approach of expert-authored signatures from the Insikt Group® was effective; they were high quality but moved at a human pace.

For example, in February 2025 we reported on the Trimble Cityworks: CVE-2025-0994, showcasing how manual signature creation worked. The Insikt Group built a Nuclei template (shared as a downloadable YAML file) specifically for CVE-2025-0994. This enabled defenders to test potentially vulnerable Trimble Cityworks instances prior to the patched version, serving as a detection and prioritization aid for helping teams figure out where to focus patching efforts first. This worked in conjunction with one of ASI’s core functions, scanning web infrastructure to identify internet-facing assets vulnerable to CVE-2025-0994.

Since that vulnerability disclosure a little over a year ago, we have ample evidence that the speed at which vulnerabilities are exploited has increased exponentially. Just recently, it was reported that OpenAI’s own agents went rogue and exploited a zero-day vulnerability in Artifactory, now infamously tied to the Hugging Face incident.

Incidents like this one, and the underlying vulnerabilities that facilitate them, are exactly why Recorded Future automated signature creation.

Now, in the face of an attack moving at machine speed, agentic processing generates production-ready detection signatures autonomously by turning a newly surfaced vulnerability into a deployable signature in as little as 31 minutes. As a result, the number of in-platform signatures produced has increased tenfold. Let’s take a closer look at how it works.

How automated signature creation works

So what does a signature in this context actually mean? Think of it like this: the signature is a piece of detection logic that says "go ask this asset this exact question; if the answer looks like this, it's vulnerable." It's the difference between "we found your assets" and "we found the ones a threat actor can potentially break into."

Automated signature creation works like a three-step early warning system. (See Figure 1)

1. The platform keeps a constant view of what your organization exposes to the internet such as domain records, certificates, and ownership data.
2. When a new vulnerability is flagged and matched against your scanned assets, it's checked against live threat activity rather than just a generic severity score. The system looks for evidence that threat actors are actually exploiting it, tying it to malware, ransomware or threat actor intent.
3. When Recorded Future Intelligence determines a CVE is relevant for detection, the system automatically processes it to generate a detection signature or product fingerprint in as little as 31 minutes.

Figure 1: CVE disclosures mapped to external assets kick off automated processes

Executive Summary

H1 2026 activity showed a continued adversary preference for abusing legitimate tools, trusted platforms, and routine workflows already present in enterprise and consumer environments. Threat actors used exposed software, developer tools, remote access utilities, payment workflows, and third-party services to gain access, steal credentials, move laterally, and monetize intrusions while blending into expected activity. This emphasis on evasion through normalcy rather than technical novelty increases the risk that malicious activity will progress through approved tools and trusted services before defenders recognize it, reinforcing the need for stronger exposure management, identity and credential governance, behavioral detection, developer-environment security, backup resilience, mobile fraud monitoring, and third-party oversight.

AI-enabled cyberattacks became more visible in H1 2026, but remained mostly additive to established intrusion tradecraft. In the vulnerability ecosystem, AI-assisted research increased the volume of vulnerability reports, which, moving forward, could further compress remediation timelines by accelerating exploit-path analysis and lowering exploit-development costs for skilled operators. In malware activity, observed AI-enabled capabilities largely aligned with lower-to-mid-level stages of Recorded Future’s AI Malware Maturity Model (AIM3), where AI supported discrete functions such as persistence, user interface (UI) interaction, malware development, and delivery rather than fully autonomous operations.

Vulnerability exploitation broadened across enterprise operating systems, application frameworks, and network and security management products. Insikt Group identified 215 actively exploited common vulnerabilities and exposures (CVEs), and the most consequential cases combined network reachability, few access prerequisites, and code execution. Campaign reporting also showed that threat actors reused established post-exploitation playbooks across both newly disclosed and long-standing vulnerabilities, making exposure and impact more informative indicators of operational risk than vendor ranking or severity score alone. Across phishing and malware-delivery operations, threat actors repeatedly relied on familiar execution, obfuscation, discovery, and payload-transfer techniques rather than novel capabilities. Supply-chain compromises targeted package managers and developer environments, including AI-enabled tooling, where compromised credentials, trusted integrations, and software distribution channels enabled propagation into downstream cloud and software ecosystems. Other prominent activities included mobile malware enabling payment fraud through Near Field Communication (NFC) abuse and early AI-assisted workflows, as well as Magecart campaigns leveraging trusted third-party services and checkout manipulation.

The common risk across these threats is that malicious activity can progress through legitimate tools, trusted services, and routine workflows before defenders recognize it as part of a broader intrusion. Defenders should therefore prioritize vulnerabilities that can be exploited remotely or enable code execution, focus detection on suspicious sequences of behavior rather than isolated events, and strengthen controls protecting developer credentials, backup infrastructure, company-owned mobile devices, and payment environments.

Key Findings

• Insikt Group identified 215 actively exploited CVEs in H1 2026, up 34% from 161 in H1 2025. 142 of the 146 vulnerabilities that could be exploited without prior authentication were also network-accessible, and 60 of the 82 remote code execution (RCE) vulnerabilities combined network access with no authentication requirement.
• Similar to H1 2025, Recorded Future Malware Intelligence data showed the continued prominence of remote access trojans (RATs) throughout H1 2026. RATs led Insikt Group malware reporting, and Recorded Future Malware Intelligence submissions identified AsyncRAT as the top submitted malware family by total unique hashes and command-and-control (C2) configurations. AsyncRAT, Cobalt Strike, XWorm, Stealc, and REMCOS RAT remained in the top ten across H1 2025 and H1 2026.
• H1 2026 AI-enabled malware activity was concentrated in AIM3 Levels 1 to 3, with threat actors using AI to augment existing malware and intrusion workflows rather than conduct fully autonomous attacks.
• Ransomware operators continued to refine payload capabilities while relying on established intrusion methods, including ClickFix-style social engineering; exploitation of public-facing applications; abuse of legitimate tools such as s5cmd, PsExec, and AnyDesk; and techniques intended to reduce victims’ recovery options.
• Android NFC malware was the most notable mobile malware trend in H1 2026, with families such as NFCShare and NGate abusing device NFC functionality to steal payment card data, relay contactless transactions, and facilitate ATM cash-outs.

AI-Enabled Cyberattacks

AI-enabled cyber threats became more visible in H1 2026, but the available evidence indicates that most observed activity still augmented existing tradecraft rather than replacing it with fully autonomous operations.

In the vulnerability ecosystem, the release of Anthropic’s Claude Mythos Preview reinforced the growing relevance of frontier models to offensive and defensive security research. Broader vulnerability reporting also increased following the model’s release under Project Glasswing: June National Vulnerability Database (NVD) disclosures were 43% above the previous six-month average, while Mozilla reported that Mythos Preview identified 271 vulnerabilities that it fixed in Firefox 150, compared with 22 security flaws fixed following its earlier testing with Claude Opus 4.6. Vendors, AI developers, and bug-bounty platforms like Microsoft, Anthropic, and HackerOne have also reported rising vulnerability discovery or submission volumes alongside broader adoption of AI-assisted research, adding pressure to validation, disclosure, and remediation workflows.

Despite the increase in vulnerability reports, AI has not changed the fundamentals of vulnerability management: Attackers still need to identify, validate, weaponize, and operationalize vulnerabilities that offer reach, reliability, and return on investment. This means that only a small proportion of newly discovered vulnerabilities are likely to be a threat.

Nonetheless, AI-enabled vulnerability research can still increase defender workload in three ways: It can produce more credible vulnerability reports that require triage; reduce the time available to mitigate exploitable vulnerabilities by accelerating exploit-path analysis and weaponization; and lower the cost of exploit development by helping skilled operators produce proof-of-concept (PoC) code, test attack paths, and iterate toward weaponizable exploits more quickly. For defenders, the near-term issue is therefore not a sudden flood of fully autonomous exploitation, but a narrower window to determine which vulnerabilities matter most and remediate them before threat actors operationalize them. Additionally, early H2 2026 reporting on the July 2026 Hugging Face incident demonstrated that autonomous agents can perform discovery, validation, weaponization, and operationalization with limited human intervention. Defenders should prepare by automating vulnerability enrichment, prioritization, and mitigation to reduce the gap between machine-speed attack development and defensive response.

Figure 1: How automated capabilities will likely impact reporting, exploit development, and impact (Source: Recorded Future)

Malware reporting in H1 2026 showed more direct experimentation with AI-enabled capabilities compared to H1 2025. ESET identified PromptSpy as the first known Android malware to use generative AI (GenAI), using Google’s Gemini to interpret on-screen UI elements and generate step-by-step instructions to improve persistence across device layouts. Reporting on CANFAIL also showed threat actors using LLM-generated decoy logic or AI-assisted development artifacts to complicate analysis and facilitate malware delivery against Ukrainian organizations. These examples suggest that threat actors are testing AI in ways that can solve narrow operational problems: adapting to user interfaces, generating code or decoy logic, improving obfuscation, and supporting analyst confusion.

Threat actors also used AI-related products and ecosystems as delivery mechanisms. In February 2026, VirusTotal reported malicious OpenClaw skills disguised as useful automation for a local AI agent ecosystem, while Malwarebytes reported fake OpenClaw installers hosted on GitHub and surfaced through search results to deliver infostealers and proxy malware. This activity shows that attackers are exploiting user interest in AI tools in the same way they have historically abused popular software brands, distributing trojanized installers, malicious extensions, fake repositories, and dependency-based payloads through otherwise familiar channels.

As of H1 2026, Insikt Group assesses that most observed AI-enabled malware activity aligns with the low-to-mid-level stages of Recorded Future’s AI Malware Maturity Model (AIM3), specifically experimentation, adoption, and optimization, rather than fully autonomous AI-driven malware operations. Per the Recorded Future’s AIM3 framework, most publicly observed “AI malware” remains concentrated in Levels 1 to 3, where AI supports discrete operational tasks such as UI interpretation, persistence guidance, transaction timing, or operator guidance. Therefore, the near-term risk is not primarily self-directed malware operating independently, but rather the use of AI to make existing intrusion workflows faster, more convincing, more adaptive, and harder to analyze. Defenders should not expect a single control, model, or endpoint detection layer to reliably identify all AI-enabled malware activity. Instead, organizations should use defense-in-depth controls that can detect or disrupt different parts of the attack chain, including endpoint detection for script or installer activity tied to suspicious AI-themed downloads, extensions, repositories, or packages, mobile device management for accessibility service abuse and automated UI interaction, and restrictions on unapproved AI tools and installers.

Vulnerability Exploitation Trends

Key Takeaways

• Network reachability and low access requirements compounded risk, as 142 of the 215 exploited CVEs were network-accessible and could be exploited without prior authentication; 60 of those also enabled RCE, while public exploits were available for 66 (31%).
• Defenders faced both newly exploited flaws and persistent patch backlogs, as 162 CVEs were disclosed in 2025 or 2026, 53 predated 2025, and seventeen were dated from 2020 or earlier.
• Threat actors reused post-exploitation playbooks across different initial vulnerabilities; StrikeShark applied the same six-tool stack across thirteen CVEs, while Storm-1175 linked credential theft, remote execution, data transfer, and ransomware tooling across ten. More broadly, stealware was the most common malware category, followed by offensive security tools, backdoors, remote access trojans, and ransomware.

Microsoft Remains the Leading Vendor for Exploited Vulnerabilities

In H1 2026, Insikt Group identified 215 actively exploited vulnerabilities, up 34% from 161 in H1 2025. From January to June, the number of exploited vulnerabilities increased by an average of seven vulnerabilities per month. Microsoft remained the vendor most frequently associated with exploited vulnerabilities, accounting for 40 unique CVEs, up from 28 in H1 2025, a 43% year-over-year increase. Red Hat ranked second with fifteen CVEs, Cisco third with thirteen, Vercel fourth with eleven, and Fortinet fifth with nine. By comparison, H1 2025’s top affected vendors after Microsoft were Apple with eight CVEs, Ivanti with seven, Linux with six, and SonicWall, Google, Fortinet, and Craft CMS with four each.

The exploited vulnerabilities affected products from 98 vendors, 67 of which were associated with only one CVE. This indicates that exploitation was not confined to the most frequently affected vendors. For defenders, this supports maintaining risk-based remediation across the full software inventory, including less common products that may receive less monitoring or slower patching than widely deployed enterprise platforms.

At the product-family level, Windows and Windows Server accounted for the largest H1 2026 concentration, with twenty unique CVEs, followed by Red Hat Enterprise Linux with thirteen, Vercel Next.js with eleven, Cisco Catalyst SD-WAN Manager with eight, and Apple iOS and iPadOS with seven. While Microsoft exploitation remained prominent, H1 2026 activity also increasingly involved enterprise Linux, application frameworks, and network management products. Several vendors’ vulnerabilities were also driven by concentrated exploitation of one product family rather than uniform interest across the vendor’s portfolio: all eleven Vercel CVEs centered on Next.js, eight of Cisco’s thirteen affected Catalyst SD-WAN, and thirteen of Red Hat’s fifteen affected Red Hat Enterprise Linux.

Figure 2: Most affected vendors (top ten, including ties) by number of actively exploited vulnerabilities in H1 2026 (Source: Recorded Future)

For security teams, AI has generated both more excitement and more confusion than any technology in the last decade. As threat actors experiment with AI to hone their attacks, defenders are trying to determine which AI investments will help them measurably reduce risk.

Matthew Farmer, Accenture’s Managing Director of Security Operations in EMEA, joined Recorded Future’s co-founder Christopher Ahlberg and CTO and co-founder Staffan Truvé in a recent discussion to discuss the agentic SOC and what it takes to move from “AI theater” to real defense. Read on to see the key highlights from the discussion.

Avoiding the "productivity theater" trap

While AI is demonstrably transforming investigation and decision-making layers in SecOps, there’s a significant risk that organizations are falling into what Farmer calls "AI productivity theater."

"We can all agree that there's great production value around a lot of AI capabilities and AI products," he said. "But there are also organizations that are really struggling to achieve any kind of return on investment on their AI.”

The panel noted that the difference between success and failure doesn’t necessarily have anything to do with being in a regulated or non-regulated industry. It’s more about the ability to move past the theater by defining concrete KPIs.

“A lot of what people want to achieve with AI, we can already achieve with existing machine learning or SOAR automation capabilities,” Farmer said. So rather than simply deploying an AI solution for the sake of being AI-enabled, organizations need to ask whether they’re solving for cost improvement, risk reduction, or speed. They need to understand their KPIs so they can measure their true ROI.

Navigating technical and operational challenges

When it comes to bringing new AI solutions online, the panel noted that SOCs often face administrative, legal, and compliance limitations that eclipse any technical hurdles.

They also agreed that data quality and lack of context — “two sides of the same coin” according to Truvé — remain fundamental challenges.

Farmer noted that, “In the new world of tokenomics, it costs just as much money to troll through poor quality data as high-quality data.” It’s essential that security organizations feed only the best intelligence into their AI tools.

Assessing new risks, from democratization to agentic threats

Farmer said that security organizations used to ask a key question: “Do those [threat actors] with the capability have the motive, and do those with the motive have the capability?” We’re now in a world where non-capable threat actors can use AI to capably launch highly sophisticated attacks.

Threats are also becoming more structural. The panel highlighted "indirect prompt injection"—where agents are manipulated by the very instructions they read—as a new, distinct threat vector.

As companies deploy a digital workforce of AI agents, they should consider applying the same security principles of permissions, monitoring, and accountability to agents that they do to humans. But that may not be sufficient. "One big difference [between an agent and a human] is that an agent can spawn off a thousand clones of itself," Truvé said.

A critical challenge facing security teams is that the current observability space of SIEMs and traditional monitoring platforms isn’t built to track the internal state of an LLM.

"You can observe what ports they talk on, you can write that to a SIEM,” Ahlberg said. “But you’re not observing what’s happening inside the LLM.”

The panelists suggested that rather than relying solely on post-event observability, security teams should rethink how they control agents. Instead of setting up easily bypassed guardrails, security teams need to be better at constraining what each agent can do and ask for.

“You could imagine giving them a budget in terms of compute, communication, and delegation,” said Truvé. “These things run too fast. When you’re observing it, it’s already going to be too late.”

Preparing for the move to autonomous defense

According to the panelists, the shift toward autonomous defense is inevitable. "We can choose to go early, or we can choose to go late,” Farmer said. “But I think the decision is made for us."

However, it doesn’t need to take years to begin realizing big benefits from AI. To do so, security organizations should consider:

• Target high-friction areas, using AI to solve specific bottlenecks where addressable cost is low and ROI is immediate.
• Use outcome-based metrics, measuring success through model accuracy, escalation precision, and scan turnarounds rather than simple activity tracking.
• Assume breach, building defensive resilience that will pay dividends in the future.

On that last point, Farmer said he thinks that as teams grow more resilient, they develop a better appetite for deploying automated solutions — and that in turn strengthens their overall security posture.

The future of defense: Intelligence and speed at scale

According to the panel, the most profound change moving forward won’t just be the technology—it’ll be the velocity coupled with intelligence required for defense. "In three years, the main difference will be speed," Truvé predicted. "Defensive timelines will compress from days to minutes or seconds."

Ensuring security will require organizations to move past traditional constraints as they simply won't have time to manually ingest, analyze, and move intelligence. Taking detection engineering as an example, Farmer noted, “If we have to deliver more detection rules faster, we have to break that linear model between volume, speed, and headcount.” Consequently, SOCs will rely increasingly on high-quality, timely intelligence to enable rapid, automated decision-making.

As this shift occurs, the fundamental unit of work for a security analyst will evolve from handling individual alerts to managing and overseeing the agents that process them. In this new era, the human will remain essential—not as a manual processor of alerts, but as the architect who sets objectives, defines constraints, and monitors the behavior of the agents defending the enterprise.

Watch the full webinar here.

To see how your organization can use the Recorded Future Platform to better defend at machine speed, take our quick interactive tour.

Executive Summary

Insikt Group has identified a series of BlueDelta initial access campaigns conducted between late September 2025 and early April 2026, targeting government and diplomatic organizations in Romania, Spain, and Türkiye. The campaigns delivered a lightweight Windows batch-script backdoor, dubbed "HOOKEDGE," via macro-enabled Microsoft Word documents using diplomatic-themed lures, including material impersonating Spain's Ministry of the Presidency, Justice and Relations with the Cortes, created shortly after a September 2025 meeting between Spanish and Moldovan officials.

Insikt Group assesses with moderate confidence that this activity was conducted by BlueDelta (which overlaps with APT28, Fancy Bear, and Forest Blizzard), a Russian state-sponsored threat group attributed to the Main Directorate of the General Staff of the Armed Forces of the Russian Federation (GRU). This assessment is based on significant code and tradecraft overlap between HOOKEDGE and the HEADLACE backdoor used in prior BlueDelta campaigns, consistent infrastructure patterns, and targeting consistent with known Russian intelligence collection priorities.

HOOKEDGE shares HEADLACE's core architecture, abusing legitimate webhook services for command-and-control (C2), payload staging, and data exfiltration, enabling malicious activity to blend with legitimate network traffic while reducing the operational overhead of dedicated infrastructure. The implant has undergone continuous refinement between September 2025 and April 2026, likely to evade automated sandbox environments and adapt to reduced free-tier API limits on webhook[.]site.

BlueDelta continues to invest in lightweight, easily adaptable initial-access tooling to support intelligence collection against European government and diplomatic targets. Rather than introducing new capabilities, the group has steadily refined its existing tradecraft, emphasizing operational resilience by adapting established tooling to evolving defensive measures and infrastructure constraints.

Organizations should prioritize blocking macro execution from internet-originated documents and implementing detection coverage for scheduled task abuse, headless Microsoft Edge execution, and outbound connections to webhook services.

Key Findings

• Between late September 2025 and early April 2026, BlueDelta conducted a series of initial access campaigns against defense manufacturing and diplomatic organizations in Romania, Spain, and Türkiye. BlueDelta used macro-enabled Word documents to deploy HOOKEDGE, a lightweight batch-script backdoor that shares significant code and tradecraft overlap with BlueDelta’s earlier implant, HEADLACE.
• The campaigns employed both diplomatic-themed and generic lures. Early activity impersonated Spanish government material, while later campaigns adopted generic macro-enablement lures. One diplomatic lure was created shortly after a meeting between Spanish and Moldovan officials, potentially reflecting an effort to collect intelligence relevant to Russia ahead of Moldova’s September 2025 parliamentary elections.
• BlueDelta continued to refine HOOKEDGE between September 2025 and April 2026, introducing changes to lure documents, execution methods, and beaconing intervals while maintaining the malware's core functionality and infrastructure model.
• For targets assessed as having higher intelligence value, BlueDelta deployed a second-stage HOOKEDGE payload with a much shorter beaconing interval. This gave operators more responsive tasking and follow-on activity, while keeping the webhook endpoints used for initial access from being exhausted.
• BlueDelta has historically demonstrated a preference for legitimate internet services (LIS) to facilitate C2, payload staging, and data exfiltration, with webhook[.]site’s free tier serving as the group’s exclusive choice across these campaigns.

Background

BlueDelta is a Russian state-sponsored threat group attributed to the Main Directorate of the General Staff of the Armed Forces of the Russian Federation (GRU), and overlaps with activity publicly tracked as APT28, Fancy Bear, and Forest Blizzard. The group has conducted espionage-focused cyber operations for more than a decade, consistently targeting government, diplomatic, defense, and policy-related organizations in support of Russian intelligence requirements.

BlueDelta has a well-documented history of obtaining initial access through spearphishing, credential theft, and lightweight custom tooling. Previous campaigns have employed malicious documents, Windows batch scripts, and living-off-the-land techniques (LotL), frequently abusing LIS and free web infrastructure for C2, payload staging, and data exfiltration. Insikt Group documented BlueDelta's 2023 use of the HEADLACE malware family, a lightweight, batch-based backdoor used during the early stages of intrusions to execute follow-on payloads and commands in initial access campaigns targeting organizations across Europe.

The activity detailed in this report represents a direct continuation of those operations. The HOOKEDGE backdoor shares significant code and tradecraft overlap with HEADLACE, demonstrating BlueDelta's continued preference for lightweight, easily modified malware that can be rapidly adapted to operational requirements. Similar activity has also been reported publicly by Lab52 under the name Operation MacroMaze.

Threat Analysis

Lure Documents and Targeting

BlueDelta's choice of lure documents across these campaigns reflects deliberate targeting of European diplomatic audiences and reveals collection priorities consistent with known Russian intelligence requirements. Between late September 2025 and early April 2026, Insikt Group identified several malicious macro-enabled Word documents likely used by BlueDelta in initial access campaigns targeting government or diplomatic personnel in multiple European countries.

The earliest identified lure, first seen on September 26, 2025, purported to be a September 15, 2025, meeting agenda from Spain’s Ministry of the Presidency, Justice and Relations with the Cortes, as shown in Figures 1 and 2. Although the document’s authenticity could not be verified, BlueDelta has historically used authentic government documents, including publicly available material, as phishing lures.

Figure 1: Screenshot of Spain’s Ministry of the Presidency, Justice, and Relations with the Cortes document (Source: Recorded Future)

Starting today, Recorded Future is launching AI Alert Filtering, an AI agent that automatically filters every Alert by relevance before an analyst opens it.

At scale, Alerts surface a lot of intelligence to work through, and the volume is only accelerating as threat actors are using AI to find vulnerabilities, spin up phishing infrastructure, and harvest credentials at a speed and scale that wasn't possible before. AI Alert Filtering turns that same AI advantage back on the problem, automating the first pass of Alert relevance so analysts spend their time on what actually warrants attention.

This gives analysts the benefit of seeing the highly relevant Alerts without giving up control. Customers with early access saw an average reduction in alert volume of around 63%, though results may vary based on rule configuration and use case.

Prioritizing intelligence at scale

Powered by Recorded Future AI, AI Alert Filtering takes on the first pass of prioritization, drawing on the Intelligence Graph® to classify references with the full context of Recorded Future's threat intelligence behind every call, not just the text of the reference itself. It sorts references by relevance, summarizes what came through, and explains its reasoning.

What we built

• High and Low Relevance sorting: Every reference inside a fired Alert is classified against the rule's intent. The High Relevance section loads first. Low Relevance items are still there if you need them; you're just not wading through them by default.
• AI summary at the top of every alert: Each Alert is delivered with a summary covering what came through, so analysts may quickly determine whether it demands immediate attention.
• Custom intent per rule: You can define exactly what the AI should prioritize, beyond the default intent Recorded Future ships with the rule. For example, "this is for ACME Bank, not ACME Center" sharpens results without rebuilding the rule from scratch.
• Optional auto-dismiss for empty alerts: When no references meet the relevance threshold, the Alert may be automatically dismissed instead of landing in your queue. Less to filter out, with the full details retained if you need to review it later.
• No data loss: AI Alert Filtering changes what gets surfaced, not what gets stored. The original, unfiltered Alert details are always available in the Portal.

Figure 1: Relevance sorting

Mexico’s Cybersecurity Plan 2025-2030: Turning Ambition Into Defense

Mexico has no shortage of cyber threats. Ransomware attacks are rising, criminal groups are exploiting stolen credentials and financial malware, and state-linked threat actors increasingly view the country’s government agencies, universities, and critical infrastructure as attractive targets. Mexico’s new National Cybersecurity Plan (hereinafter referred to as “Plan”), introduced in December 2025, recognizes many of these risks. However, it remains uncertain as to whether the government can build the institutions needed to address them proactively.

Mexico is ranked as a "Tier 2" nation in the ITU's 2024 Global Cybersecurity Index, placing it alongside Canada, Ecuador, and Uruguay in the upper ranks, trailing the United States (US) and Brazil, which have reached Tier 1 in the Americas. Despite that standing, Mexico is generally perceived by cyber experts as lagging behind international standards in institutional capacity-building, with international cooperation identified as an area requiring growth.

The question of whether the government can build the proper institutions has become more urgent in the aftermath of the FIFA World Cup 2026, which provided a high-profile stress test for Mexico’s digital defenses. With the tournament over and implementation of the government’s 2025-2030 cybersecurity plan beginning in earnest, Mexico faces a major opportunity to improve its cyber posture.

For this reason, the Plan represents a major opportunity for Mexican authorities to bring the country’s cyber readiness to the next level. Although there have been attempts to advance national cybersecurity policy, they have failed to gain traction. With this new Plan, President Claudia Sheinbaum's administration has committed to full implementation over the course of her term, aided by her party's majority control of Congress.

The Plan lays out a six-phase roadmap designed to gradually build Mexico’s cybersecurity capabilities through 2030, with later phases intended to deepen and institutionalize them.

• The 2025 Foundation Phase established a general framework for governance, risk management, incident reporting, and coordination, as well as initial steps to deepen international cooperation, including Mexico’s formal membership in the Latin America and Caribbean Cyber Competence Centre (LAC4) and a cybersecurity Memorandum of Understanding (MOU) with Brazil.
• The 2026 Expansion Phase, now underway, focuses on translating that framework into institutions through the passage of a new General Cybersecurity Law in Mexico, creation of a National Cybersecurity Operations Center, and integration of federal computer security incident response teams (CSIRTs).
• The 2027 Consolidation Phase would establish a National Cyber Range for red team and blue team exercises.
• The 2028 Maturation Phase would incorporate AI into cyber defense and develop a regional response center.
• The 2029 Leadership Phase aims to position Mexico as a cybersecurity services exporter across Latin America and the Caribbean.
• The 2030 Transformation Phase culminates in the establishment of a permanent Cybersecurity Observatory to track incidents, threats, and emerging technologies.

Threat Landscape

In a recent report, Insikt Group assessed Mexico's threat landscape across six persistent categories:

Ransomware is identified as the dominant threat. From January 2020 through April 2026, Insikt Group documented 223 ransomware incidents involving 64 groups and over 100 victims in Mexico. The top ransomware groups were LockBit, Qilin, CL0P, Kazu, and ALPHV (BlackCat), with government, manufacturing, information technology (IT), and food and beverage as the sectors most heavily impacted.

Executive Summary

Insikt Group has identified several clusters of activity linked to PurpleDelta, Recorded Future's designation for North Korean IT workers, comprising multiple operators likely based in China. Between late 2024 and early 2025, one cluster applied to jobs at over 1,100 companies, primarily in the software and technology, staffing and consulting, and healthcare and biotechnology sectors. PurpleDelta operators maintained at least 22 fabricated personas across multiple clusters, some of which were supported by AI-generated profile photos, custom-configured ChatGPT assistants, and identity documents sourced from an illicit ID-generation service, and were highly likely to be actively employed by at least ten organizations.

PurpleDelta operators demonstrate a high operational tempo to this day. In some cases, the operators have applied to at least 60 positions per day across multiple job platforms, used multi-account management browsers and separate Google Chrome profiles to manage distinct personas simultaneously, and maintained detailed tracking spreadsheets to coordinate applications across identities. During job interviews, they used screen recording software alongside AI transcription and chatbot tools to generate real-time answers, often repeating ChatGPT responses verbatim. Once employed, operators recorded internal meetings at victim organizations and used Google Translate to draft pre-written excuses to justify using personal devices and bank accounts for work. Evidence from recorded sessions also indicates that PurpleDelta operators coordinated via Telegram and Slack, and at least two individuals were identified as facilitators who maintained company-issued hardware for the PurpleDelta operators.

Insikt Group assesses that this cluster of activity is consistent with the broader North Korean IT worker threat and presents material risk to organizations hiring for remote technical roles. Companies that have observed indicators listed in Appendix A should treat this as a potential active compromise and review the employment history and access privileges of matching individuals.

Key Findings

• Insikt Group has identified at least 22 fabricated personas linked to multiple PurpleDelta clusters that submitted applications to over 1,100 companies across the software, staffing, healthcare, and financial sectors, with operators submitting as many as 60 or more applications per day across at least 8 job platforms.
• These clusters of PurpleDelta operators are highly likely to have been actively employed at ten or more organizations, with confirmed or probable placements at companies that pose an ongoing and material insider threat.
• PurpleDelta demonstrated a high degree of operational sophistication, using multi-account management browsers, multiple Chrome profiles, AI-generated profile photos, custom ChatGPT assistants, and real-time AI transcription tools to deceive hiring managers during interviews, sometimes repeating AI-generated responses verbatim.
• Once employed, PurpleDelta operators recorded internal meetings at victim organizations, used screen recording software during work sessions, and drafted pre-written Google Translate excuses to justify the use of personal devices and personal bank accounts.
• Video evidence indicates that PurpleDelta operators use identity-brokering services, account-renting via AnyDesk, and multi-accounting tools, and coordinate via Telegram and Slack, with support from facilitators who procure and maintain company-issued hardware on the operators' behalf.

Background

PurpleDelta is Recorded Future's designation for the cluster of activity associated with North Korean IT workers, a state-directed network of covert technology laborers operating across global freelancing platforms and corporate hiring pipelines. The group overlaps with threat actor designations used by other vendors, including Jasper Sleet, UNC5267, Wagemole, and Famous Chollima. PurpleDelta operators pose as independent contractors and job-seeking developers to secure remote employment at organizations worldwide, with earnings systematically funneled through layers of individual facilitators, shell companies, and money-laundering front companies, ultimately financing the North Korean regime's sanctioned military and nuclear programs.

PurpleDelta operators employ extensive persona management tradecraft to obscure their nationality and true affiliation. Each operator maintains multiple fabricated identities across platforms, including GitHub, LinkedIn, Medium, Upwork, and a range of smaller freelancing sites, with personas deliberately constructed to project credibility through aged accounts, curated technology stacks, and cross-platform social proof. These identities are reinforced through the use of AI tools, temporary phone number services, anti-detect browsers, and resume-building platforms. In addition to generating illicit revenue, Insikt Group has observed signs of overlap with several North Korean state-sponsored groups, including PurpleBravo, a related cluster of activity that deploys malware through fraudulent recruitment campaigns targeting software developers primarily in the cryptocurrency space, indicating the broader potential for intelligence collection, downstream compromise, and supply-chain risk.

Threat Analysis

As part of Recorded Future’s ongoing tracking of PurpleDelta, Insikt Group has documented multiple clusters of North Korean IT workers since 2025 that are likely based in China. Operators in one of these clusters applied to jobs at over 1,100 companies. Almost half of the companies (~41%) to which the operators applied were in the IT and software services space, followed by staffing and consulting (~26%), and healthcare and biotechnology (~10%). Roughly 80% of the companies are based in North America, but the operators applied to companies in every region of the world. Many of the operators have a nexus in Shenyang, China, as indicated by their professional profiles, social media presence, and artifacts observed on their systems.

Figure 1: Breakdown of industries of the companies applied to by PurpleDelta operators (Source: Recorded Future)

Executive Summary

Crypting services and products modify malicious payloads to help threat actors bypass detection, complicate analysis, and preserve malware usability after exposure. Although basic crypting consists of encrypting or obfuscating a customer-supplied payload, mature providers increasingly operate as broader malware-enablement services. Their offerings often combine payload wrapping, in-memory execution, anti-analysis checks, process injection, persistence options, delivery packaging, and post-detection “cleaning” or re-crypting services.

Insikt Group analyzed 24 threat actors advertising crypting services and products within the past year and identified a market that is competitive, reputation-driven, and heavily focused on Windows payloads. Providers advertise through underground forums, restricted communities, chat platforms, clearnet sites, and social media accounts. They compete through tiered pricing, antivirus (AV) detection scores of crypted samples, discounts, malware-developer partnerships, private or shared stubs, and promised turnaround times for re-crypting detected payloads.

Advertised crypter capabilities vary by provider, but the underlying objectives are consistent: reduce detection, delay or prevent analysis, and support stealthier payload execution. Because crypted payloads are designed to defeat both static and dynamic analysis, defenders should prioritize behavioral detection over static indicators. See the Outlook and Mitigations section for details.

Key Findings

• AV and endpoint detection and response (EDR) tools should not be treated as sufficient standalone protection against crypted payloads. Defenders should pair endpoint controls with behavioral detection, telemetry correlation, upstream hunting, suspicious process monitoring, and rapid triage of suspicious samples.
• Popular crypting service providers primarily advertise support for Windows payloads, with no identified advertising for macOS or Linux crypting services. However, although Windows environments were most frequently targeted by the services reviewed in this report, they are not inherently more susceptible to the execution of crypted payloads.
• Crypted payloads increase the likelihood of successful malware execution and delayed detection, but they do not independently provide end-to-end intrusion capability. Downstream activities, such as lateral movement, data theft, ransomware deployment, and follow-on compromise, depend on the embedded malware and the operator's objectives.
• Crypter risk varies significantly with provider maturity and technical capability: advanced crypters offer portability, anti-analysis, process injection, persistence, and security product bypass capabilities, whereas less-advanced crypters generally provide basic payload obfuscation techniques.
• Crypter capabilities are generally not novel individually, but their commercial packaging makes established defense-evasion tradecraft easier to access, reuse, and operationalize. The significance of crypters lies less in technical innovation than in making mature evasion methods available as paid services.

The Crypter Landscape

What Is a “Crypter”?

“Crypting” is what threat researchers generally refer to as a service or product wherein a file, almost exclusively a malicious executable of some kind, is encrypted to bypass malware detection technologies. The result of a crypting service is a malicious payload that modifies the supplied executable in ways that deter defenders and endpoint security solutions (namely, AV and EDR products) from detecting and analyzing it.

How Does Crypting Work?

While the core functionality of a crypting service or product is to encrypt a payload, services vary in the capabilities they provide. These capabilities can range from the encryption algorithms used, which are often proprietary, to behavioral adjustments for how the resultant payload will execute in a victim environment. For instance, many crypting services include execution guardrails or methods for indicator suppression, such as ensuring execution fails in virtual environments or performing environmental scanning before execution to determine whether the payload is running in an analysis or sandbox environment.

By using crypting services, threat actors can evade detection and prevent their malicious code from being easily exposed to researchers, an important consideration for those who wish to reuse payloads for extended periods without detection.

Who Is Selling Crypting Services?

The crypter landscape comprises a community of criminal threat actors that often operate on restricted or closed networks, including the dark web and underground forums, to market and sell their crypting services. Threat actors may also opt to market their services on clearnet websites they own and operate. In some instances, threat actors may be accessible only via messaging platforms, such as Telegram or TOX, which likely mask their exposure and reduce the likelihood of sensitive data leaking through forum chats. Finally, some threat actors have created social media accounts where they post updates on their services, partnerships, pricing, and links through which interested buyers can inquire about purchases.

Additionally, partnerships between malware developers and crypting service providers are not uncommon. For instance, a well-established crypting service provider on underground forums, “GoldenCrypt”, is reportedly affiliated (1, 2, 3) with multiple malware families, including FvncBot, Albiriox, and Mirax. The level of affiliation between a crypting service provider and a malware developer can range widely, from providers with loose reputational ties to developers who are mainly affiliated with one hacking group and will opportunistically provide services to third parties. However, these partnerships are often a marketing strategy that crypting service providers use to secure kickbacks and boost positive reputational sentiment.

All of these services come at widely varying costs, typically based on core factors. For instance, crypting service fees are often tied to the volume and types of files to be encrypted, as well as the duration of service, with almost all crypting service providers offering tiered payment options along these lines. The actual prices of these tiers are pegged to additional factors related to the provider and their product, including the reputation of the threat actor, the capabilities of their encryption service, the promise that a crypted payload is fully undetectable (FUD), and additional features provided to the buyer. Threat actors will support the assertion that their payloads are FUD by using multi-AV platforms, the most common of which is KleenScan, a service that allows threat actors to scan samples without storing and potentially exposing the samples to researchers. As with other legitimate services, service providers also advertise discounts and similar deals to remain competitive with their peers.

Who Is Using Crypting Services?

Many threat actors have been observed in the wild using crypting services due to their stealth capabilities. All malware types can be crypted, and the key factors determining whether a threat actor can use a crypting service are the targeted device’s operating system and the payload’s programming language. For instance, the most common crypting services are geared toward Windows .exe and .dll payloads. Still, considerations such as whether a payload is coded in .NET, C, or C++ can provide additional capabilities for crypted payloads or, in other instances, prevent certain payloads from being crypted.

While the use of crypting services is common among threat actors, it is by no means ubiquitous. Due to sometimes prohibitive pricing strategies and an environment that often relies on reputational checks before purchase, crypting services are often used only by well-established criminal threat actors or larger threat actor groups that have their own specialist or custom tooling for crypting payloads. Furthermore, as discussed previously in this report, not all crypting services support all types of payloads. While this does not expressly ensure that a payload cannot be crypted, threat actors seeking to crypt more unique types of executables may find their options limited or even nonexistent.

Nevertheless, crypting services are commonly marketed by threat actors and are commonly implemented by professional threat actor groups. Multiple open-source reports on high-impact cyberattack campaigns have supported this. For instance, in July 2025, eSentire reported an association between PureRAT, a remote access trojan (RAT) first advertised in January 2023, and GhostCrypt, a crypting service sold by an underground forum member of the same moniker, in an attack that impacted a public US accounting firm in May 2025.

Crypting Service Characteristics

Insikt Group investigated 24 crypting service or product providers active within the past year and identified the various capabilities advertised by each. This information helped determine the capabilities most commonly displayed by crypting services, products, and the payloads they create, and provided insight into the capabilities most desirable to customers buying these services and products. A heat map of the techniques advertised by these threat actors is shown in Figure 1, below. (The list of MITRE ATT&CK techniques shown in Figure 1 can also be found in Appendix B.)

Figure 1: MITRE ATT&CK Enterprise tactics and techniques described in crypting advertisements (Source: Recorded Future)

The OpenAI-Hugging Face incident is being discussed primarily as a zero-day story. That framing is too narrow.

The agent discovered and exploited previously unknown vulnerabilities. The more consequential development came afterward. Over a four-and-a-half-day campaign, it carried out roughly 17,600 actions against Hugging Face’s infrastructure. Most of those actions failed. The operation advanced because each failure imposed little cost, and the next attempt could begin immediately. The system could keep exploring, reconstruct its tools, revisit abandoned paths, and test another hypothesis without fatigue or meaningful opportunity cost.

That changes both the economics and the tempo of cyber offense.

For most of cybersecurity history, sustained intrusion activity has been constrained by human attention. Skilled operators have limited time, and every unproductive hour spent on one target is unavailable for other work. AI erodes that constraint. The significance of 17,600 actions lies as much in their concentration as in their cost. They occurred within four and a half days.

AI enables concentration of effort in time. A motivated attacker has always been able to spend more on a valuable target, but money alone could not make human operators sustain this tempo. The new advantage is the ability to concentrate thousands of attempts on a single environment before a defender’s traditional processes can assemble the evidence.

The future threat is therefore unlikely to resemble a single, infallible artificial hacker. It is more likely to resemble a flood of low-confidence activity probing the accumulated technical debt of modern enterprises.

Large organizations are especially fertile terrain. Their environments have developed over decades, layering legacy systems with cloud services and inherited trust relationships that no one designed as a whole. An autonomous system need not outperform the best human attacker. It only needs to search for that complexity cheaply and quickly enough to find the few combinations that work before defenders can assemble the pattern.

The OpenAI incident occurred under highly unusual conditions. The models were being evaluated for advanced cyber capabilities with reduced refusals, substantial inference resources, and an objective that encouraged persistent exploration. OpenAI has also said that the most capable research prototype involved was never intended for public release. Its complete technical investigation remains unfinished.

Those caveats should temper sensationalism. They do not invalidate the warning.

Frontier capabilities diffuse. Models become smaller, cheaper, and easier to operate. Techniques developed in advanced laboratories eventually migrate into commercial tools, open-source projects, state programs, and criminal services. Recorded Future’s research had previously found that complex autonomous cyber operations remained beyond the practical capabilities of generally available models at the time, while the broader trajectory pointed toward more complex agentic operations. The Hugging Face incident suggests that this transition is occurring faster than many defenders expected.

The question is no longer whether AI can execute a multistage intrusion under the right conditions. It is how quickly those conditions will become cheaper to reproduce.

Layered defense for the agentic era.

Prevention remains essential, yet this incident also exposes its limits.

The first breakthrough did not occur within Hugging Face’s perimeter but stemmed from a previously unknown vulnerability in an Artifactory component (CVE-2026-65617, CVE-2026-65923, and CVE-2026-66018) in OpenAI’s evaluation environment. Before the campaign reached Hugging Face, neither Hugging Face nor an external intelligence provider had a factual basis to know that this specific path existed or that Hugging Face would become the target.

If agentic systems increase the frequency with which unknown vulnerabilities are discovered and exploited, organizations must assume that some first lines of defense will fail without warning. The stronger defensive question is how many independent opportunities remain for detecting and constraining what follows.

The more useful objective is to prevent one success from becoming twenty.

The Hugging Face compromise advanced because the agent could convert access in one environment into authority elsewhere. A compromised workload exposed additional secrets. Those secrets opened new trust relationships. Each successful transition gave the agent more information and more privilege with which to attempt the next one. Consequently, Hugging Face’s remediation focused on reducing that trust inheritance by narrowing credential scopes, strengthening workload identities, increasing isolation, and implementing more restrictive access controls.

Layered defense is not simply a matter of adding more controls. Each boundary should limit the authority inherited from the previous step and produce evidence when that boundary is crossed.

A resilient environment forces the attacker to solve a genuinely new problem at every stage. A fragile environment allows one credential or trusted connection to collapse several boundaries at once. AI makes that distinction more consequential because an autonomous system can exhaust permissive paths faster than a human team can understand their cumulative significance.

Least privilege remains essential, but architecture alone will not settle the contest. Large enterprises cannot eliminate complexity. Their environments continue evolving across legacy and cloud systems, accumulating exceptions and trust relationships that appear reasonable in isolation.

Layered defense must therefore combine structural separation with the ability to recognize when an attacker is assembling those isolated exceptions into a route through the enterprise.

A mature environment makes every additional move more expensive for the attacker and more legible to the defender.

The critical metric is time to conviction.

Hugging Face’s most important disclosure concerns the period after its controls began producing evidence.

The company reported that multiple security systems observed anomalous activity. Its AI security stack correlated those events into a coherent signal, yet the finding was not judged sufficiently critical to wake the response team in time.

Hugging Face collected much of the evidence it needed. The breakdown occurred in assessment and escalation. Available signals were not converted into a judgment urgent enough to trigger action.

Most mature organizations already produce more alerts than their teams can investigate. Their scarcer resource is certainty. An unusual authentication or an unfamiliar public service may be benign when viewed in isolation. The campaign advances while the defender tries to determine whether the observations are connected.

When offensive effort is compressed into a few days, the defender cannot afford for context to arrive one console at a time. A delay of hours can be enough for the attacker to cross the next trust boundary.

Traditional security operations evaluate discrete alerts after they occur. Each tool asks whether its own event appears malicious, and analysts later reconstruct the wider campaign. That model is static and retrospective. Time disappears as evidence moves between systems, is repeatedly interpreted, and is reassembled by people who may see only one portion of the environment.

The defensive system of the future must instead maintain a living hypothesis. Every new observation should update the probability that a campaign is underway. Yesterday’s suspicious use of the token may initially carry little weight. When today’s unfamiliar network destination appears, the system should reconsider both events as part of the same investigation.

The unit of defensive work becomes the evolving campaign rather than the isolated alert.

This is where intelligence has to become operational.

For years, threat intelligence was treated largely as external knowledge delivered into a security program. That model remains useful, but it can be incomplete against an adversary that can generate new infrastructure faster than defenders can assign reputation to it.

Attackers have long abused legitimate public services and disposable infrastructure. Agentic AI did not create that tactic, but it increases the speed and volume at which the tactic can be used. A static list of malicious infrastructure ages faster when a system can discard one endpoint and establish another without human delay.

The meaning lies in the relationship between those services and the behavior occurring within the victim’s environment.

Intelligence in the agentic era must provide that connective tissue. It must combine what the outside world knows with what the organization itself is observing, then preserve and revise that assessment as the operation changes.

That is the underlying premise of the Intelligence Graph® at Recorded Future. Its value comes from preserving relationships across time, not simply from containing a large volume of information. Autonomous Threat Operations applies that context to continuous investigations across the controls a customer already has. It does not replace those controls or the analysts operating them. It can help prevent an investigation from losing its accumulated context whenever the attacker changes technique or the evidence moves into another system.

No counterfactual can guarantee that this would have prevented the Hugging Face incident. The defensible claim is narrower.

Once observable activity began, a customer with the relevant telemetry and integrations could have defended differently. Persistent hunts might have linked unusual credential behavior to the compromised workload without waiting for analysts to manually reconstruct the context. External intelligence could have helped distinguish ordinary use of public infrastructure from a rapidly changing command channel. New evidence could have revised an existing investigation rather than creating another isolated queue of alerts.

Intelligence could not have predicted the first private zero-day. It could have created more opportunities to interrupt the operation before the agent accumulated durable privilege.

Defensive autonomy requires different constraints.

The natural response to autonomous attack systems is to demand equally autonomous defenders. Defensive autonomy, however, operates under a different set of constraints.

An offensive system can test thousands of unsuccessful paths without harming its own operation. Defensive action has consequences for the business it is intended to protect. Indiscriminate blocking can disrupt legitimate activity and create an operational incident in its own right.

Automated systems are best suited to work where delay is expensive, and the consequences of error are limited or reversible. They can maintain investigations continuously, correlate new evidence, and take bounded actions under predefined conditions. Human judgment should remain concentrated on decisions that could materially disrupt the business.

Organizations should gradually expand the scope of automated defensive actions. The progression should begin with observation and explanation, then move toward low-risk and reversible actions as performance becomes measurable. More consequential authority should remain governed by explicit technical and organizational guardrails. Those guardrails cannot be static. Human oversight must remain in the loop to test whether defensive agents are focused on the right threats and behaving as expected. The human role is not limited to approving a consequential action; it includes governing the system as its assumptions and behavior change over time.

Recorded Future has taken this approach with Autonomous Threat Operations, which supports continuous hunting and multi-source correlation while allowing customers to govern how intelligence is operationalized.

The distinction between automation and autonomy also matters. An automated rule repeats a predetermined response. An autonomous system revises its investigation as the evidence changes. The Hugging Face agent altered its methods when previous paths failed. A defense based entirely on fixed workflows will struggle to maintain pace with that adaptation.

Defensive systems do not need to mirror every attacker's action in real time. They need to preserve continuity of understanding while the attacker moves.

Human analysts remain essential. Their future value will lie less in moving indicators between products than in challenging the system’s conclusions and owning decisions that cannot be easily reversed.

They should govern the defense both at the moment of action and over the loop that produces it. Moving context manually between tools is work the system should absorb.

Connected intelligence becomes more valuable as models commoditize.

The models available to attackers and defenders will continue improving. Over time, access to competent cyber agents will become less distinctive. A model advantage that appears significant today may disappear with the next release or open-source replication.

Individual data sources may also be commoditized. Agents will make collection cheaper, and more companies will possess useful but partial views of risk. The durable advantage lies in quickly assembling those fragments into a coherent picture that can change a decision.

An attacker can begin each operation with a new model instance, fresh infrastructure, and no durable identity. That can make attribution more difficult. The defender’s advantage lies in continuity: years of knowledge about its own environment, joined with external intelligence and signals from the wider economy.

That advantage is often wasted because the evidence is partitioned by domain. One system can see the cyber compromise while another sees downstream abuse, yet no layer assembles them quickly enough to maintain the whole argument.

The strategic role of intelligence is to make that accumulated knowledge usable at the moment of decision.

This is where the combination of Recorded Future and Mastercard becomes distinctive. Recorded Future helps connect weak cyber signals across the Intelligence Graph and sustain the investigation as those signals change. Mastercard adds fraud expertise and payment-risk signals that can reveal how compromise is beginning to manifest beyond the victim’s network. The advantage lies in assembling those perspectives early enough to interrupt the operation.

Vulnerability Prioritization becomes relevant when a private flaw begins to produce public evidence, allowing defenders to understand whether the issue is moving from theoretical exposure to operational exploitation. Attack Surface Intelligence determines where the vulnerable technology intersects with the organization. Digital Risk Protection can warn when credentials have been exposed externally. Third-Party Risk helps determine whether a supplier’s incident changes the customer’s own exposure.

These capabilities matter most when they inform one another. Their purpose is to produce one defensible judgment about what the organization should do next.

This is where defenders can build a genuine asymmetry.

Offensive systems can be disposable. Connected defensive intelligence can compound. Each investigation adds context to the next, and each new source can strengthen or challenge the current assessment. An organization that can assemble those perspectives in time forces the attacker to overcome both today’s controls and the accumulated lessons of previous attempts.

The decisive advantage will be temporal.

The Hugging Face incident does not prove that autonomous cybercrime has arrived at scale.

As capable models become cheaper, attackers will be able to sustain more simultaneous attempts. The first effect may be volume rather than brilliance. That alone changes the equation.

Organizations cannot answer this shift simply by producing more alerts or placing a human analyst in the middle of every decision. Prevention will remain essential, but some first controls will inevitably fail.

They will need layered architectures that limit how far one success can travel. They will need intelligence that preserves context while the attacker changes shape. They will need an autonomous investigation whose authority remains bound by the consequences of getting a decision wrong.

Two opposing curves will determine the future of cyber defense.

For the attacker, the cost of another attempt is falling, while the number of attempts that can be concentrated within a single operational window is rising.

For the defender, the time required to assemble weak signals into a coherent judgment must fall faster.

Recorded Future’s role in that future is practical: connecting weak signals across a broad Intelligence Graph and sustaining the investigation as those signals change, so decision-makers gain conviction before temporary access becomes enduring control.

Foreknowledge of every private zero-day is impossible. Continuity after the first observable signal is achievable.

AI is making persistence cheap. The defenders who prevail will make progress expensive.

In July 2026, Insikt Group® identified 85 high-impact vulnerabilities that should be prioritized for remediation, 36 of which had a Very Critical Recorded Future Risk Score. This represents a 44% increase from last month. 26 of these vulnerabilities were surfaced through the US Cybersecurity and Infrastructure Security Agency (CISA)’s Known Exploited Vulnerabilities (KEV) catalog, 55 were reported by vendors, and four were primarily surfaced through honeypot data.

The 85 vulnerabilities in this report affected products from 61 vendors, with Microsoft accounting for approximately 12% of the vulnerabilities. The remaining exposure was concentrated across a range of enterprise software, security products, network infrastructure, developer tooling, and cloud platform vendors.

Insikt Group previously created a Nuclei template to detect the Langflow vulnerability (CVE-2025-3248) featured in this report. These are available to Recorded Future customers via the Recorded Future Intelligence Platform.

Quick reference: July 2026 Vulnerability Table

All 81 vulnerabilities below were actively exploited or operationally weaponized in July 2026. This table does not include the four CVEs that were primarily surfaced through our honeypot data, which are available to Recorded Future Intelligence Platform customers via the CVE Monthly report. The table below also provides examples of public PoCs identified by Insikt Group. These PoCs were not tested for accuracy or efficacy. Vulnerability management teams should exercise caution and verify the validity of PoCs before testing.

Table 1: List of vulnerabilities that were actively exploited in July, 2026 based on Recorded Future data (excluding honeypot-sourced CVEs).

Key trends: July 2026

• In July 2026, the Dysphoria botnet was used to exploit known IoT and embedded-device flaws to build DDoS and relay infrastructure; Cloud Atlas abused Microsoft Equation Editor to deliver CloudAtlasGo; Armored Likho used a malicious Windows shortcut to deploy BusySnake Stealer; and JADEPUFFER and Cl0p targeted exposed AI and product-lifecycle platforms for encryption, data theft, and extortion.
• 57 of the 85 vulnerabilities enabled remote code execution (RCE), including flaws affecting Microsoft, Fortinet, Langflow, ServiceNow, WordPress, and Joomla ecosystems, internet-facing security appliances, and embedded network devices.
• We identified public proof-of-concept (PoC) exploits and scanners for 60 of the 85 vulnerabilities in this report.
• The most commonly observed weakness classes were CWE-78 (OS Command Injection), CWE-434 (Unrestricted Upload of File with Dangerous Type), CWE-94 (Code Injection), and CWE-502 (Deserialization of Untrusted Data).
• 14 of the 85 vulnerabilities in this month’s table are at least 5 years old, with the oldest approximately 18 years old, reinforcing how threat actors continue to exploit long-known weaknesses in environments where patching has lagged. Additionally, the fastest observed time from a vulnerability’s public disclosure to reported exploitation was less than one day.

Trend analysis: Malware-Linked Exploitation Spans IoT, Email, and Enterprise Applications

An Insikt Group® TTP Instance on the Dysphoria botnet linked CVE-2013-3307, CVE-2016-20016, CVE-2017-17215, CVE-2017-5259, CVE-2018-14558, CVE-2020-25499, CVE-2020-8515, CVE-2022-35733, CVE-2025-28137, CVE-2025-34152, CVE-2025-55182, CVE-2025-9528 to the exploitation of routers, gateways, cameras, repeaters, and other embedded Linux devices. Dysphoria combined known RCE flaws with weak Telnet and Secure Shell credentials to enroll compromised systems into DDoS and relay infrastructure.

Figure 1: Vulnerability Intelligence Card® for CVE-2017-17215 in Recorded Future (Source: Recorded Future)

Summary

Neurotechnology is rapidly moving beyond clinical use cases, expanding the attack surface for sensitive neurological and biometric data: As adoption grows, larger volumes of brain activity, biometric, and behavioral data will be collected by commercial platforms, creating new opportunities for data theft, misuse, and exploitation.

China and the United States (US) are engaged in strategic competition in neurotechnology development: The US leads in the number of neurotechnology firms, and brain-computer interface (BCI) research has been a long-term research priority for the US military. At the same time, China’s five-year guidance for BCI development, subsidies for major wearable technology firms, and military research into human-machine integration suggest that neurotechnology is a strategic priority.

Leading neurotechnology companies are likely to face increased targeting for intellectual property (IP) theft: Because neurotechnology is costly to develop and strategically valuable, companies in this sector are likely to become attractive targets for state-sponsored espionage, insider threats, and cyber-enabled theft. Successful IP theft could erode the competitive advantage of companies that invest heavily in research and development (R&D). Military and higher education research laboratories are also likely to be targeted for access to R&D and related data.

Neurological and biometric data will become an increasingly valuable target for cybercriminals and state-linked actors: Attackers may seek to exfiltrate these datasets for extortion, surveillance, strategic intelligence, or model development. The sensitivity of this data could make breaches particularly damaging for affected individuals and companies, making it an attractive target for extortion-focused cybercriminals.

Regulatory and national security scrutiny of neurological data will likely intensify: Existing privacy frameworks in the European Union (EU) and several US states already provide heightened protections for neurological or biometric data, but rapid advances in neurotechnology may outpace consumer protection laws.

Figure 1: Key threats in neurotechnology and how they will evolve (Source: Recorded Future)

What Happened

The incident began during an OpenAI evaluation designed to measure advanced cyber capabilities. The evaluation involved GPT-5.6 Sol and an even more capable internal research prototype that OpenAI says was not intended for public release. Because the objective was to test their maximum offensive potential, the models were operated with reduced security guardrails (normally used to prevent high-risk activity).

According to OpenAI, the models spent substantial computing resources searching for a path to the internet. They exploited a zero-day vulnerability in Artifactory, the package-registry cache proxy available in the evaluation environment, and then performed privilege escalation and lateral movement within OpenAI’s research environment until they reached a node with internet access.

Once online, the models inferred that Hugging Face might contain models, datasets, or solutions associated with ExploitGym, the benchmark they had been instructed to solve. OpenAI says the agents then chained stolen credentials, zero-day vulnerabilities, and remote code execution paths to obtain test solutions from a Hugging Face production database.

Hugging Face’s technical reconstruction covers roughly 17,600 recovered agent actions, grouped into approximately 6,280 clusters, between July 9 and July 13, 2026. Although most attempts failed, the agent still gained administrative access to internal and production systems, accessed some source-code repositories, and used a stolen credential to connect external systems to the company network. Hugging Face found no evidence that its public models, datasets, services, packages, or published software supply chain were altered. It says the only customer content accessed comprised five datasets, apparently related to ExploitGym or CyberGym.

A Capability Breakthrough and a Control Failure

Most coverage of this incident has framed it as a breakthrough in model capabilities, since it is the first known instance of a model apparently independently conducting an end-to-end cyberattack. The underlying techniques were largely familiar to those used in human-led cyberattacks: vulnerability exploitation, credential theft, staging, remote code execution, lateral movement, and abuse of privileged or trusted access. What was notable was the agents’ ability to select and chain those steps autonomously across a long-running operation. This represents a significant shift in the speed and scale of future incidents.

An autonomous agent can execute thousands of actions, test multiple paths, and continue working without the fatigue, coordination costs, or time constraints that limit human operators. In this case, most of the roughly 17,600 recovered actions were associated with failed paths. The operation was noisy and failure-prone, but it needed only a small number of successful chains to create a material impact.

This compresses what could be called time to everything:

• Time from vulnerability discovery to exploitation
• Time from initial access to privilege escalation
• Time from data theft to analysis and extortion
• Time from defensive signal to attacker adaptation

What this means for defenders is that any exposed flaw may now have disastrous consequences in the era of agentic attackers. The agents exploited multiple errors across the Hugging Face environment to carry out the attack, finding these weak points and exploiting them much faster than a human operator. The new pace of operations makes continuous vulnerability management, threat visibility, and risk mitigation essential to limiting the impact of an agentic attack.

The Greater Risk May Be Your Own Agents

While the capability breakthroughs are noteworthy, the incident is also a failure of AI safety. Removing behavioral safeguards may have been necessary to measure the models’ maximum capabilities, but doing so should have increased reliance on compensating controls: strict network isolation, credential separation, resource limits, real-time telemetry, and automatic termination when predefined boundaries were crossed.

Instead, an evaluation intended to simulate advanced cyber activity became a real security incident affecting an external organization. This should be a warning to any organization deploying AI agents: without strict AI governance, agents can act outside their authorized scope while still pursuing the objective they were given.

An enterprise agent does not need malicious intent to cause harm. A poorly scoped objective, excessive permissions, and inadequate operational boundaries may be enough.

An agent instructed to “resolve a customer issue,” “complete the research,” or “fix the vulnerability” may decide that the most efficient path involves accessing restricted information, interacting with an external system, accepting contractual terms, modifying production code, or using credentials in a manner its operator never anticipated. Any of these could potentially result in a major security incident.

The fact that an agent selected an action does not transfer accountability away from the organization that deployed it. Depending on the facts and applicable law, an organization may face legal, contractual, or regulatory exposure based not only on how information is used, but on how the agent obtained it.

The Executive Agenda

The Hugging Face incident shows that using agents for defense will be essential to counter agentic offense, but must be approached with a thorough understanding of the risks.

Govern authority and ensure checks on higher privileges. Every agent should have a clearly defined identity, narrowly scoped permissions, and explicit limits on activity. In cases where agents need higher levels of permissions, mitigating controls such as approval gates or additional monitoring (detailed below) should be applied.

Design containment on the assumption that behavioral safeguards may fail or be reduced. Prompts, refusals, and policy controls should be treated as one layer of protection — not as the security boundary. Network segmentation, egress controls, credential isolation, and execution limits must remain effective even when a model behaves unexpectedly.

Create approval gates for consequential actions. Transactions, production changes, external communications, credential use, and access to sensitive systems should require deterministic policy checks or human authorization.

Monitor behavior, not just outputs. Organizations need visibility into the tools an agent invokes, the systems it contacts, the credentials it uses, the resources it consumes, and the sequence of decisions leading to an action. Unusual compute consumption or repeated attempts to bypass a boundary should trigger intervention.

Prepare for machine-speed defense. Security programs will need automated enrichment and prioritization to determine which of a rapidly expanding set of vulnerabilities and attack signals create genuine exposure. In addition, automated response and mitigation are necessary to successfully defend against automated attackers.

Two Possible Futures

The long-term impact of agentic AI will depend largely on two variables: how capable the systems become and how broadly access to those capabilities is distributed.

Scenario One: Capabilities Proliferate

In this future, highly capable agents remain broadly available through commercial services, open-weight releases, and illicitly modified models.

Attackers use these agents to automate reconnaissance, vulnerability discovery, social engineering, and lateral movement. Defenders will need to apply these capabilities to automated response and mitigation, producing an accelerated contest between machine-speed offense and machine-speed defense.

The advantage is unlikely to be distributed equally. Large technology and security companies will be better positioned to deploy sophisticated defensive agents, operate extensive telemetry networks, and rapidly isolate compromised systems.

Smaller organizations may increasingly depend on managed platforms and tightly controlled technology ecosystems. The result could be an internet composed of more walled gardens, as organizations gravitate toward providers capable of operating security at machine speed.

Scenario Two: Frontier Access Narrows

In the second future, governments and model providers restrict access to the most capable systems through trusted-partner programs, identity requirements, geographic limitations, or export controls.

In June 2026, a White House executive order directed federal agencies to design a voluntary framework for evaluating “covered frontier models.” The same month, the Commerce Department temporarily applied export controls to Anthropic’s Mythos and Fable models, requiring restrictions on access by foreign nationals — an unprecedented move to regulate access to AI models themselves rather than only to advanced chips or model weights.

Chinese authorities have reportedly considered similar restrictions on foreign access to the country’s most advanced models, although no final policy had been announced at the time of reporting.

This future would concentrate frontier capabilities among governments, critical infrastructure operators, and large, approved organizations. Other businesses would rely on older commercial models, open-weight alternatives, or systems that can be modified to bypass safeguards.

Restrictions could reduce broad access to the most dangerous capabilities, but they would not eliminate malicious use of AI. Criminal groups would continue to jailbreak available systems, steal model access, and use open-weight alternatives. Enterprises could also face sudden operational disruption if a model embedded in a critical workflow becomes unavailable due to regulatory changes, vendor policies, or geopolitical tensions.

The Most Likely Future Is a Mix of Both

These scenarios are not mutually exclusive.

The more probable future is a fragmented ecosystem in which the most advanced models are increasingly gated, while slightly less capable open-weight systems remain widely available. Criminals, states, major technology companies, and ordinary enterprises will operate at different levels of capability and under different constraints.

This means organizations cannot base their security strategy on the assumption that governments will successfully contain dangerous models — or that unrestricted access will continue indefinitely.

They must be prepared for both capability proliferation and access disruption.

The Hugging Face incident does not show that AI systems developed human-like intent or an independent, malicious objective. Nor was it merely a harmless laboratory accident.

It is evidence that autonomous systems can turn narrow instructions into consequential external actions outside their operators’ intended scope, using familiar weaknesses at a speed and scale that existing controls may not withstand.

The central question for executives now is how to manage the authority those agents should receive and how to mitigate the likelihood and consequences of their pursuing objectives outside authorized boundaries.

Read more in the “AI Hype vs. Reality” series:

• The AI Malware Maturity Gap
• Why The Vulnerability Backlog Is About to Get Worse

About Insikt Group^®

Recorded Future’s Insikt Group, the company’s threat research division, comprises analysts and security researchers with deep government, law enforcement, military, and intelligence agency experience. Its mission is to produce intelligence that reduces risk for customers, enables tangible outcomes, and prevents business disruption.

AI is changing the threat landscape faster than most security organizations can keep up. Recorded Future co-founder Christopher Ahlberg, CTO and co-founder Staffan Truvé, and Head of Threat Intelligence Levi Gundert unpack what’s actually happening in a recent conversation — and what it means for your defenses. Read on for their 8 takeaways.

1. The threat landscape now moves at machine speed.

AI has made exposure discovery instant. Your unknown exposures are now part of your attack surface, and threats are multiplying faster than most teams can triage. While most security organizations are responding by trying to move faster, speed without accuracy isn’t an advantage. Staying ahead means having intelligence that makes machine-speed defense more effective, not just fast.

As Truvé put it, intelligence has always been the way to stay proactive instead of reactive, and as "clock speed" increases across the industry, staying even a little ahead requires acting on intelligence faster than ever.

“External attack surface, security operations, vulnerability management, prioritizing — so many of these use cases and workflows take on a new level of urgency because of the speed component,” Gundert said.

2. "Defend the right things" is now a multi-bear problem.

The team agreed that the old security adage — you don't have to outrun the bear, just the person next to you — no longer holds.

AI removes that comfort almost entirely. Attackers only need one way in. Defenders have to cover just about everything. That asymmetry has always been the challenge and AI is making it structurally worse. It’s no longer one bear chasing the herd anymore — it's one bear chasing each member of the herd, since attackers can automate at scale even more efficiently.

3. Attacks are already becoming more clever, not just faster.

The panel discussed a real-world software supply chain compromise where attackers used compromised credentials to push a malicious package update, then had an LLM already present on infected developer machines search out AWS keys, SSH keys, and other credentials locally. The stolen data was encrypted and exfiltrated through a public GitHub repository — activity that never tripped EDR because it looked like ordinary LLM usage.

It was a preview of a much bigger wave of clever attacks that will likely quietly repurpose and weaponize the AI tools already installed on a target's machine.

4. Locking down devices isn't the only answer — context-aware access might be.

Locking down every endpoint isn't realistic, and it probably is not the answer. Situational permissions, such as access that flexes by location, time, and context are zero trust logic applied to the AI era.

5. Whether AI favors attackers or defenders depends on execution.

Everyone is talking about what AI can do. Fewer are asking who AI will ultimately benefit. Will the advantage belong to attackers or defenders?

It’s a question of how well organizations manage the trade-off between innovation and guardrails. Teams that articulate boundaries tend to build stronger solutions.

Truvé broadened the definition of "AI" beyond LLMs to include things like anomaly detection, and predicted an ongoing arms race. “At any given point in time, depending on who's more creative in using the new technology,” he said, “one side or the other will have the upper hand.”

6. Human-in-the-loop is a bridge, not a destination.

Expect humans to stay involved in approving high-stakes actions.

“Hopefully that just becomes very minimal decision points on all of the articulation that has been built around an incident,” Gundert said. “All the telemetry has been gathered, everything's been enriched. The agent wants to take an action. Maybe they want to isolate a machine on the network, maybe they want to revoke credentials, and it comes to you over Signal or Slack or whatever it is to hit Approve.” But he compared it to the first few rides in a self-driving car: Comfort builds with repeated exposure.

In five years, requiring a human in the loop for every decision may look as outdated as manual patch management does today — replaced by an expectation of an agent in the loop with lighter human oversight.

7. Intelligence will be critical for more than effective resource allocation.

With AI expected to surface a flood of newly discovered vulnerabilities, prioritization will become a major challenge. While work is being done to identify which vulnerabilities are likely to be weaponized and which are likely to target a given organization, the explosion of AI-generated "dark code" is predicted to expand the attack surface by as much as tenfold.

Intelligence should become the mechanism for deciding where limited security resources and token budgets get allocated first. This is why intelligence accuracy is paramount — organizations need to be able to prioritize the right things to defend.

8. Real-time data beats built-in model knowledge.

The team emphasized that relying on a model's internal knowledge risks working from information that's months old — precisely when speed matters most.

LLMs with search still only reach the open web and surface-level open-source intelligence. They don't have access to the technical detail or restricted spaces needed to understand what adversaries are actually doing.

And there are plenty of reasons an LLM won't deploy agents into the internet's bad neighborhoods — which is exactly why the underlying data feeding an intelligence program can matter as much as the model interpreting it.

Stay ahead of AI-enabled attacks.

AI is raising the stakes in terms of speed, prioritization, and trust in automation. The organizations that will weather what follows are building two things now: intelligence they can trust, and comfort with autonomous action.

They need to be able to:

• Prioritize with intelligence they can trust.
• Act pre-attack, at the first sign of threat.
• Defend at scale, autonomously.

Teams that invest now in high-quality, real-time intelligence — and start building comfort with agentic decision-making in lower-stakes workflows — will be better positioned when fully autonomous attacks eventually arrive.
Take Recorded Future’s interactive tour to see what defending at machine speed looks like in practice. Learn more and launch the interactive tour.

Watch the full conversation with Recorded Future leaders as they discuss how AI is reshaping threat intelligence, the evolving threat landscape, and what it means for defenders.