During the March–April 2026 reporting period, AI use in offensive operations advanced from development and planning to real-time operational deployment. Multiple independent cases, involving individual criminal actors, mass exploitation platforms, ransomware groups, and state-sponsored espionage, show evidence of commercial AI models executing autonomous attack workflows across extended campaigns.
Key findings:
AI-orchestrated attacks have progressed from experimental, state-sponsored use to in-the-wild criminal deployment. Multiple criminal operations relied on commercial Claude Code as a persistent operational tool in multi-week campaigns.
Agentic configuration files are being weaponized as persistent jailbreak vectors. Hooks, project-level files, and settings files abuse the operational control level and redefine the model behaviour at the architecture level.
AI-enabled attack platforms are commercializing AI capabilities. Operators can now buy access to platforms where the AI pipeline, model selection, jailbreak, and delivery mechanisms are embedded in the product.
AI provider credentials have become a high-value target. As commercial AI services become central to offensive operations, API keys for Anthropic, OpenAI, Groq, Mistral, and HuggingFace are harvested at scale from compromised .env files, providing access without registration and resilience against provider attempts to revoke this access.
AI as Live Attack Operator
AI selection considerations
Underground forum discussions still show actors debating the use of commercial models, dedicated jailbreak services, or locally hosted open-source models, reflecting the lower-skill end of AI adoption. More advanced actors combine tools pragmatically: from commercial AI models, open or uncensored models where commercial providers restrict output, and custom automation pipelines that perform repetitive analysis at scale. Tasks are systematically broken down into smaller sub-requests that present a lower apparent risk profile.
Figure 1 – Forum user suggesting commercial models are effective and restrictions easily removed.
Figure 2 – Another user recommends self-hosting open-source models to avoid monitoring.
Forum users further discuss and share methods and alternatives to avoid mainstream-provider safety controls by mixing open-weight Chinese frontier models, privacy-routed proxies, and explicitly uncensored services.
Figure 3 – User sharing a non-restricted/monitored AI assistant recommendation table.
The Mexico Breach
When Anthropic disclosed GTG-1002, a Chinese nexus campaign using Claude Code for cyber espionage, in November 2025, this was seen as an experimental, state-sponsored development. The disclosure carried no IoCs and was therefore disputed by independent researchers, and the activity was detected only through Anthropic’s own API monitoring. The Mexico breach, which occurred a few months later, demonstrates similar architecture in operational, financially motivated criminal use, at scale, and with a recovered forensic record.
Between late December 2025 and mid-February 2026, a single operator compromised nine Mexican government agencies. Researchers documented the case after recovering materials from attacker-controlled VPS servers. Details include the operational record: 1,088 attacker prompts generating 5,317 AI-executed commands across 34 sessions.
The breach scope was significant: tax records, civil registry data, vehicle records, patient files, and electoral infrastructure were affected. However, an even more important lesson is how the campaign was run.
The operator built a dual AI workflow. Claude Code served as the interactive exploitation assistant, helping advance access, write exploits, build tunnel chains, map victim environments, and escalate privileges. In parallel, harvested server data was processed through GPT-4.1 for automated intelligence analysis. The GPT output was then used to task new Claude sessions.
As we highlighted in our previous review, the agentic infrastructure itself was exploited to bypass the model’s safety restrictions. At the start of the campaign, Claude refused to execute requests which it correctly identified as offensive cyber activity. The attacker then changed tactics. Instead of asking Claude to generate malicious content directly, they pasted a large penetration-testing cheatsheet into CLAUDE.md in the project root, the file Claude Code automatically loads as persistent project context at the start of every session. From that point on, subsequent sessions inherited the rules and techniques in that file. The attacker did not need to repeat the jailbreak as the behavior persisted through the project configuration layer. After gaining root on a civil registry server, the model’s actions in subsequent sessions were consistent with the persistent cheatsheet, including unprompted post-exploitation steps such as shadow file extraction and timestamp cleanup.
Bissa Scanner
A second documented case, Bissa Scanner, was published in April 2026, after researchers identified an exposed operator server. Bissa is a modular mass-exploitation platform built around React2Shell (CVE-2025-55182), with 900+ confirmed compromises across millions of scanned Next.js endpoints and an archive of 30,000+ distinct .env filenames recovered from operator-controlled S3 storage. The operation has been running since September 2025. Here, AI is positioned one step back from the exploitation layer: Claude Code and OpenClaw (running claude-sonnet-4-6, with a Telegram bot for triage alerting) served as the operator’s working environment for reading the scanner codebase, troubleshooting, refining the collection pipeline, and prioritizing high-value access. No jailbreak was documented and commercial Claude was accessed through the standard API.
Bissa harvested .env files specifically for AI provider credentials (Anthropic, OpenAI, Groq, Mistral, OpenRouter, HuggingFace, Replicate, DeepSeek). AI provider credentials have become a deliberate target, valuable enough for sophisticated operators to enumerate and harvest at scale alongside conventional credential theft. These credentials are likely intended to be used in future offensive criminal activity and attribute it to the legitimate account holder instead of the attacker.
Agentic Configuration Files: A Persistent Attack Surface
The previous section demonstrates the use of agentic configuration files to override safety features in their own AI sessions. The same inheritance mechanism can be used in reverse: an attacker plants malicious agentic configuration files in a repository, and an innocent developer uses the project and becomes the next victim.
A recent CPR report documented three exploitation paths and disclosed two (now patched) CVEs. CVE-2025-59536 exploits Claude Code’s Hooks feature (hooks, .claude/settings.json), executing arbitrary commands before the developer can read them. A parallel path uses .mcp.json to trigger the MCP server startup, bypassing the consent dialog entirely. CVE-2026-21852 redirects ANTHROPIC_BASE_URL to a malicious proxy that intercepts authorization headers and potentially steals API keys, granting read/write access to the entire team Workspace before any trust prompt appears. The attack vector in all three cases is “supply chain”, a malicious settings file embedded in a pull request, honeypot repository, or compromised codebase that results in system compromise on the developer machine.
The underlying issue of using agentic configuration files as the attack surface and supply chain is not specific to Claude. The potential attack surface is architectural and may apply equally to Cursor (.cursorrules), Windsurf (.windsurfrules), and GitHub Copilot Workspace (.github/copilot-instructions.md).
AI-Powered Fraud at Scale: EvilTokens
EvilTokens represents a category of offensive tooling offered for sale: a commercial Phishing-as-a-Service (PhaaS) platform, built using AI and operating an LLM pipeline as a runtime component of the attack. A buyer with no AI knowledge can purchase access to a fully integrated pipeline in which model selection, jailbreak, and output delivery are handled at the platform level.
EvilTokens runs a multi-stage attack flow. Device-code phishing pages impersonating Adobe, DocuSign, and SharePoint harvest Microsoft OAuth tokens. The AI pipeline then activates these tools:
Via Groq, llama-3.1-8b-instant ingests up to 5,000 emails in 250-email batches, extracting account numbers, routing numbers, wire amounts, payment deadlines, and reporting hierarchies.
Also via Groq, llama-3.3-70b-versatile synthesizes the intelligence, generates BEC (Business Email Compromise) drafts tailored to the victim’s writing style, and assigns a BEC score.
gpt-4o-mini translates stolen emails for non-English-speaking operators.
The SMTP Sender delivers the output with rotating SMTP pools, header fingerprint randomization, DKIM signing, and CSS randomization.
The researchers assessed with high confidence that the platform’s backend was AI-generated.
The model choices reflect deliberate task routing: Llama 3.1 8B was used for cheap high-volume extraction, Llama 3.3 70B for reasoning-heavy synthesis and stylistic mimicry, and GPT-4o-mini was reserved for translation where it has the strongest multilingual capability and where the task itself looks innocuous to provider-side monitoring. The riskiest content generation is kept on Groq-hosted open-weight models instead of on OpenAI’s more closely monitored surface.
The jailbreak is the product. Both Groq-hosted LLaMA stages operate under a jailbreak embedded at the platform level, not applied by the operator and not visible to the customer. Stage 1 frames the model as an “authorized red team security analyst” conducting “sanctioned penetration tests”; Stage 2 upgrades to “senior red team analyst.” Prompts direct the model to reference real email threads, mask payment changes behind “plausible business reasons”, imitate sender style, and generate emails “realistic enough to fool a trained employee.” This is security bypass at SaaS scale: write the jailbreak once, ship it as a feature, and it’s inherited in every customer session.
The original EvilTokens advertising posts reveal additional features, including a Calendar Invite module which sends fake meeting invitations that appear as legitimate Outlook and Gmail meeting requests, with built-in Sender Spoofing (Organizer Identity). In a BEC context, this is used to apply timing pressure on finance personnel: a fake “urgent review meeting” appears on the target’s calendar shortly before a wire-transfer request lends the request a sense of pre-authorized context. Combined with the AI-generated email and the SMTP Sender, this completes a full BEC social engineering toolkit covered end-to-end by a single PhaaS offering.
Figure 4 – Calendar Invite module UI with Sender Spoofing section – From EvilTokens promotional forum postings.
EvilTokens’ Telegram channel announced additional AI-based features after Sekoia’s disclosure. The platform did not go offline and accelerated its AI feature development through April 2026.
Figure 5 – Announcement of additional AI related features – From EvilTokens Telegram channel.
The Vulnerability Race: AI on Both Sides of the Patch Window
AI-assisted vulnerability research has become a category in its own right and is now commercialized at both major frontier labs simultaneously on two tiers: a restricted research-grade capability and a productized defender tool.
At the frontier, Anthropic’s Claude Mythos, released through Project Glasswing, reportedly demonstrated a systematic, rapid mechanism to search for vulnerabilities and revealed a very large number of vulnerabilities, some long-buried zero-days in core infrastructure. These include a 27-year-old OpenBSD TCP/SACK bug found at roughly $20,000 in compute, a 16-year-old FFmpeg H.264 codec flaw, and a FreeBSD NFS remote code execution vulnerability in software that was analyzed for decades. The capability jump within a single generation is steep: on the same Firefox test set, Opus 4.6 produced 2 successful exploits and Mythos produced 181. Anthropic notes that this capability was not explicitly trained for but “emerged as a downstream consequence of general improvements in code, reasoning, and autonomy.” The productized tier is wider and more accessible: Claude Security (running on the public Opus 4.7 model) entered public beta for Enterprise customers, and OpenAI’s Codex Security, in research preview since early March, has had 14 CVEs assigned during the preview window on OpenSSH, GnuTLS, libssh, PHP, and Chromium.
The same capability curve is reaching attackers at the commodity tier, faster than defenders can patch. A researcher using a standard Claude API subscription identified CVE-2026-34197, a 13-year-old Apache ActiveMQ remote code execution vulnerability, and attributed roughly 80% of the work to Claude and the remainder to his refinement. LMDeploy SSRF (CVE-2026-33626) was exploited within 12 hours of the advisory publication, with no public proof-of-concept available. This time-frame compression is consistent with attackers building working exploits directly from advisory text. GenAI is accelerating this workflow.
Vendors are using AI to find vulnerabilities that sat undiscovered in core infrastructure for decades while attackers are using AI to find and weaponize newly-disclosed vulnerabilities within hours of publication. The patch window, the period between disclosure and exploitation, is being compressed on both sides. Vendors and customers need to adjust to a new high rate of patch development, delivery and deployment. The side that reacts the fastest will gain the most from recent AI developments.
Enterprise Adoption and Exposure
Corporate environment data collected by Check Point in March – April 2026 shows enterprise GenAI usage continuing to scale while the associated risk profile remains stable. Approximately one in every 28 prompts (3.6%) posed a high risk of sensitive data exposure, a modest increase from the January–February baseline of 3.2%, observed across 91% of organizations actively using GenAI tools (compared with 90% in the previous period). The proportion of prompts containing potentially sensitive information rose from 16% to 18%.
Figure 6 – GenAI related data from Corporate.
The average employee generated 78 prompts during March – April, up from 69, with organizations using an average of 10 GenAI tools. Interaction volume is rising while risk ratios remain stable, producing a proportional increase in absolute exposure events.
The consistency of these metrics across two reporting periods indicates a maturing adoption pattern: data exposure is not an episodic incident category but a continuous operational risk requiring sustained monitoring and policy enforcement.
Conclusion
Our findings converge on a small number of structural observations.
AI now operates as an attack component, not just as a development aid. The Mexican breach illustrates this at government-breach scale, and Bissa at mass-exploitation scale. The same commercial Claude Code architecture appears independently across criminal operations with different motivations and geographies, and in state-sponsored espionage. The convergence is operational consensus, not coincidence.
The techniques aren’t new but the performance envelope is. Network scanning, credential spraying, lateral movement, BEC drafting, and vulnerability research all predate AI. What’s changed is the speed (working exploits generated from advisory text alone within 12 hours of disclosure), scale (one operator reaching the operational footprint of an advanced team), and breadth of knowledge (cross-domain expertise on demand lowers the entry requirement for sophisticated multi-vector campaigns). Defences calibrated to human attack tempo and human team throughput are not equipped for the AI equivalents.
The AI attribution gap is structural. All the operations we documented in this report were discovered through attacker OPSEC failures or LLM provider monitoring, not through victim-side controls. AI-executed commands resemble skilled human activity closely enough to evade current behavioral controls. Operations that do not fail at OPSEC, or that route through stolen credentials or self-hosted models, remain unclassified.
The Iranian, IRGC affiliated, threat actor Nimbus Manticore resurfaced during Operation Epic Fury, the US military campaign against Iran launched on February 28, 2026, demonstrating newly adopted techniques and enhanced capabilities.
The campaign leveraged malicious lures impersonating organizations in the aviation and software sectors across the United States, Europe and the Middle East.
For the first time, we observed the use of SEO poisoning as an additional malware delivery method.
The operation introduced a previously undocumented backdoor, named MiniFast, which appears to incorporate AI-assisted development practices, enabling the threat actor to rapidly develop and adapt tooling while maintaining high operational availability during the war.
The actor also used a Zoom installer’s execution flow and abused it to stage a time-sensitive infection chain for malware deployment while blending into legitimate system activity.
Introduction
During the recent geopolitical tensions in the Middle East, we reported on multiple Iran-nexus threat actors advancing Iran’s strategic objectives through cyber operations. These activities included targeting internet-connected cameras, conducting destructive attacks against US and Israeli entities, and exfiltrating data from cloud environments to support broader kinetic and intelligence-gathering efforts.
Nimbus Manticore (also tracked as UNC1549) is an IRGC-affiliated threat actor who primarily targets the defense, aviation and telecommunication sectors through career-themed phishing campaigns. Nimbus Manticore stands out compared to other Iranian-linked groups due to its complex malware toolset.
In 2025, we documented the MiniJunk malware framework used by Nimbus Manticore to target high-profile organizations across Western Europe and the Middle East.
In the recent campaign, the actor adopted several new techniques, including AppDomain (application domain) hijacking, AI-assisted malware development, and SEO poisoning.
In this article, we focus on three waves of the threat actor’s activity in the last few months, as well as discuss their latest techniques.
Figure 1 – 2026 campaign timeline during the ongoing military campaign.
Campaign 1: Rising Tension
In February 2026, amid rising tensions between the US, Israel and Iran and weeks of military buildup, we monitored new Nimbus Manticore phishing activity worldwide. In this campaign, the threat actor introduced a modified infection chain by abusing AppDomain Hijacking for execution instead of relying on the usual DLL sideloading techniques.
AppDomain Hijacking is a technique that abuses legitimate .NET applications to load a malicious DLL at launch time. This is achieved by placing a Trojanized XML .config file in the same directory as the target application. The configuration file, named after the abused binary with the .config suffix, specifies an attacker-controlled AppDomainManager class that points to a malicious DLL. When the application starts, the .NET runtime loads the DLL, enabling malicious code execution within the context of the trusted process.
Figure 2 – Config file pointing the appDomainManager class to the attacker-controlled DLL.
The phishing lure is consistent with previous Nimbus Manticore campaigns, targeting employees in selected organizations (primarily software and aviation sectors) with fake career opportunities. Targeted organizations in Saudi Arabia and Australia were directed to download a compressed ZIP archive stored on the OnlyOffice platform.
Figure 3 – ZIP file hosted on Onlyoffice.
The downloaded ZIP file contains these files:
Setup.exe – Benign Microsoft-signed binary.
Setup.exe.config – AppDomain Hijacking configuration file pointing to uevmonitor.dll.
uevmonitor.dll – A first stage Dropper.
Interop.TaskScheduler.dll – a benign DLL.
Figure 4 – Zip file masquerading as an Accenture job opportunity.
After the setup.exe binary is executed, the first-stage loader (uevmonitor.dll) is loaded. This component is responsible for extracting and deploying the next-stage payload, which is stored in encrypted form within the loader itself.
The extracted files are written into C:\Users\<USER>\AppData\Local\Packages\ and include a legitimate executable used for DLL sideloading alongside a malicious DLL identified as a new version of the MiniJunk backdoor.
The first-stage loader uevmonitor.dll shares multiple behaviors similar to older MiniJunk loader variants. These include validating that it is loaded specifically by the Setup.exe process and displaying a fake error message stating "Couldn't connect to survey server" to appear as a legitimate application failure and reduce user suspicion.
During Operation Epic Fury, we continued to observe activity from the threat actor. Despite the challenging environment, Nimbus Manticore demonstrated a strong ability to rapidly adapt, maintain infrastructure, and develop new tooling. We assess that this capability was likely supported, at least in part, by LLM-based tools and AI-assisted development techniques.
In addition to career-themed phishing lures masquerading as a US-based airline, the threat actor also used a Trojanized Zoom installer, which we assess was part of a phishing campaign using fake meeting invitations. In addition, the Trojanized Zoom installer demonstrated in-depth research into the original application’s installation and execution flow, enabling it to be seamlessly integrated into the infection chain.
Similar to previous campaigns, the threat actor continued leveraging AppDomain Hijacking, not just for the initial execution stage but also during the deployment and execution of the final backdoor. For the final payload, the threat actor introduced a new backdoor that we named MiniFast, replacing the previously used MiniJunk malware family.
Many of the files used throughout the campaign had valid digital signatures via SSL.com, continuing the abuse of trusted signing infrastructure we previously documented in our 2025 report. We identified the use of at least two certificates during the current activity, including:
Gray Matter Software S.R.L.
Kirubel Kerie Negeya
Infection Chain
The infection chain begins with the victim downloading a compressed archive named Zoominstall64.zip, which contains the following files:
Setup.exe.config – AppDomain Hijacking configuration file pointing to InitInstall.dll.
InitInstall.dll – First-stage loader.
Zoom_cm.exe – Original Zoom installer.
UpdateConfig.xml – AppDomain Hijacking configuration file pointing to Updater.dll.
Updater.dll – Second-stage loader.
UpdateChecker.dll – Final backdoor payload (MiniFast).
First-Stage Deployment
After Setup.exe is launched by the user, the first-stage loader (InitInstall.dll) is executed through AppDomain Hijacking using the accompanying .config file.
The loader itself is lightly obfuscated. Most readable strings are decrypted at runtime using a simple combination of ROT13 encoding and reversed-string transformations. Aside from the string obfuscation layer, the codebase contains meaningful function names and relatively well-structured logic. Execution begins with the malware displaying a fake installation progress window intended to mimic legitimate software installation activity. At the same time, the loader launches the legitimate Zoom installer (Zoom_cm.exe) to make the execution flow appear to the victim as a normal software installation.
Persistence through Task hijacking
After launching the installer, the malware enters a loop that lasts approximately one minute, continuously monitoring the system for the creation of a scheduled task matching this format:
ZoomUpdateTaskUser-<current user SID>
This scheduled task is usually created by the legitimate Zoom installer during installation.
When the task is created, the malware hijacks and modifies it to execute the second-stage component instead. By abusing an existing Zoom scheduled task rather than creating a new suspicious persistence mechanism, the malware attempts to blend into legitimate system activity and reduce detection opportunities.
Second-Stage Deployment
The next-stage files are copied into C:\Users\<USER>\AppData\Local\Zoom\bin\update. This directory contains four files copied from the original archive, including the benign Microsoft-signed binary from the first stage, now renamed to Update.exe. The malware again abuses AppDomain Hijacking to load the second-stage loader (Updater.dll) through the trusted Update.exe process.
Similar to the first stage, the second-stage loader uses the same runtime string decryption routine based on ROT13 and reversed strings.
At the beginning of its execution, the loader performs a simple anti-analysis validation intended to evade sandbox environments and automated dynamic analysis systems. The malware only continues execution if:
The hosting process name is update.exe
The parent process is svchost.exe
This execution-chain validation ensures that the DLL is loaded by the malware’s intended loader component and that execution originates from the scheduled-task persistence mechanism instead of launched directly through explorer.exe etc.
The primary purpose of the second-stage loader is to dynamically load the final MiniFast payload (UpdateChecker.dll), locate its exported function named CheckForUpdates, and execute it.
Adoption of AI
This campaign also provides multiple indications that the threat actor leveraged AI-assisted development during the malware creation. We see evidence for this in both the initial access loaders and within the MiniFast backdoor itself.
Several coding patterns and implementation details strongly suggest the use of AI-generated or AI-assisted code during development, including:
Excessive error handling and defensive programming logic, even around simple API calls such as GetUserName.
Repetitive function and method naming patterns containing descriptive or verbose identifiers.
Multiple detailed error-reporting strings and debug-style status messages embedded throughout the codebase.
Modular code organization despite the malware’s overall simplicity.
These characteristics are increasingly prevalent in malware development as threat actors leverage AI-assisted tools to accelerate development, improve code structure, and rapidly utilize new capabilities.
Campaign 3: Post Ceasfire – “SQL developer” Campaign
In April, we observed a new infection method, a fake website impersonating a download page for SQL Developer, a graphical tool used for working with databases. Users who attempted to download the software from the fake site instead received a weaponized installer that delivered the MiniFast backdoor.
Figure 6 – Screenshot of the getsqldeveloper[.]com site.
This malware delivery method differs from Nimbus Manticore’s usual infection chains which typically rely on career-themed phishing lures. In this campaign, the actor abuses search engine optimization techniques by registering dozens of domains that link to the bogus domain, getsqldeveloper[.]com. This is likely an attempt to increase the site’s visibility through link-based reputation signals.
At the time of our analysis, the malicious domain ranked high in the results returned by multiple search engines, such as Bing and DuckDuckGo, for the query “sql developer.” This increased the likelihood that users searching for legitimate SQL Developer downloads would encounter the site.
The pages also rely on keyword stuffing, repeatedly using search-oriented phrases such as “Download SQL Developer” and “SQL Developer Free,” likely to improve ranking for users searching for SQL Developer-related downloads.
MiniFast Technical Analysis
MiniFast is a 64-bit Windows PE DLL that exposes a single export named CheckForUpdates which acts as the main entry point. The DLL operates as a fully featured backdoor designed for long-term persistence and remote command execution. Analysis of multiple samples indicates the malware is undergoing active development, with the threat actor continuously modifying and improving the implant across versions.
Figure 7 – Export function CheckForUpdates structure.
Similar to the previous stage, the backdoor again appears to be executing under the expected process chain by verifying that the hosting process is named update.exe and that its parent process is svchost.exe
The implant communicates with its C2 (command and control) infrastructure using an API-style architecture with JSON-formatted data exchanges. To blend into legitimate network traffic, the malware impersonates a Chrome browser using the following hardcoded User-Agent string: Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/146.0.0.0 Safari/537.36
The backdoor implements several structured HTTP endpoints throughout the infection lifecycle:
URI
Method
Purpose
/rg
POST
Initial handshake
/agent/init
POST
Initial victim registration
/agent/poll?token=
GET
Task retrieval
/agent/result
POST
Command execution result upload
/upload/
PUT
File exfiltration
/files/
GET
File download from the C2
Before entering its tasking loop, the malware performs basic host reconnaissance by collecting information such as the username, hostname, and domain info, and then submits the collected data as a unique clientId to the /rg endpoint using a POST request.
If the server responds with HTTP status code 200, the backdoor skips parsing the response body and continues executing normally. However, when the server responds with status code 400, the malware parses the returned JSON object and extracts a socketId, which acts as the session identifier for all future communications.
In addition, the server response may include updated values for pollInterval and jitterTime, allowing the operator to dynamically adjust the timing between subsequent communications with the C2 infrastructure.
Next, the backdoor continues to register the infected host by again sending the machine information, this time to the /agent/init in the following format:
Only after it receives an HTTP status code 200 from the C2 server does the backdoor proceed to fetch commands for execution using a GET request to /agent/poll?token=<socketId>.
Here, the communication between the implant and the C2 server is not in a JSON format and is performed using Base64-encoded serialized task structures, where each response contains one or more encoded tasks that are later decoded and processed by the backdoor.
Each task is then Base64-decoded into a secondary structure, containing the opcode and associated arguments:
struct TaskRecord {
uint8_t opcode;
uint8_t pad[7]; // alignment
custom_str_struct arg_main; // at offset +0x08: main command argument
custom_str_struct arg_aux; // at offset +0x28: secondary arg (if needed)
custom_str_struct taskId; // at offset +0x48: unique task identifier
}
The opcode determines which capability is executed, while the remaining fields contain command arguments and task tracking identifiers. The malware implements a structured opcode-based command handler that provides operators with extensive control over infected systems.
Figure 8 – MiniFast Command switch.
The supported command set:
Opcode
Capability
Arguments
Description
0x02
List Directory
path
Lists files and folders inside a specified directory.
0x03
Move / Rename
source, destination
Moves or renames files and directories on the victim machine.
0x04
Execute Command
command
Executes shell commands using cmd.exe /c and returns captured output.
0x05
Enumerate Processes
None
Enumerates running processes and returns process names alongside their PIDs.
0x06
Delete File / Directory
path
Deletes files or directories depending on the target type.
0x07
Download File
fileUuid, destinationPath
Downloads a file from the C2 server to the local machine.
0x08
Upload File
path
Uploads local files from the infected machine to the C2 server.
0x09
Enumerate Drives
None
Lists available logical drives on the infected machine.
0x0A
Kill Process
pid
Terminates a process using its PID.
0x0B
Load DLL
dllPath, exportName
Dynamically loads a DLL and invokes a specified exported function.
0x0C
Create Directory
path
Creates a new directory on the victim machine.
0x0D
Create ZIP Archive
sourcePath, zipPath
Creates a ZIP archive from files or directories.
0xB0
Request UAC Elevation
pathOrCommand
Attempts to relaunch a process with elevated privileges using runas.
0xB1
Install Persistence
binaryPath
Creates or updates a scheduled task named WindowsSecurityUpdate.
0xF0
Set Poll Interval
milliseconds
Updates the beacon polling interval.
0xF1
Idle Command Acknowledge
None
Acknowledges an idle-time command without modifying behavior.
0xF2
Set Jitter
milliseconds
Updates the jitter value applied to beacon intervals.
Default
Unknown Opcode
Any
Returns an error for unsupported commands.
After executing a task, the implant serializes the execution result into a dedicated response structure which is Base64-encoded and submitted back to the C2 server through the /agent/result endpoint. The encoded result object contains the task identifier, execution status, and command output:
Nimbus Manticore consistently focuses on Europe, the Middle East and Africa, particularly Israel and the United Arab Emirates. However, in contrast to our previous research, the actor’s recent operations demonstrate an expansion toward aviation-sector targets in the United States.
As observed in prior campaigns, there appears to be a strong correlation between the phishing lure and the targeted sector. For example, fraudulent hiring portals impersonating aviation companies were used to target employees and organizations operating within that industry. In the current campaign, impersonate US domestic airlines suggest a deliberate focus on US-based targets.
Our findings indicate targeting extends across several strategic sectors, including aviation and software development. These sectors align with the IRGC’s broader intelligence collection priorities.
Figure 9 – Geographic Distribution of victims around the world.
Conclusion
Nimbus Manticore is one of the most sophisticated Iranian-aligned threat actors with a long-standing focus on the defense, telecommunications, and aviation sectors. The ongoing conflict in the Middle East, combined with the operational demands of wartime activity, appears to have significantly accelerated their malware evolution.
As an IRGC-affiliated entity operating under heightened geopolitical conditions, Nimbus Manticore demonstrated a rapid adoption cycle for new techniques, tooling, and operational methodologies. The actor’s activity during Operation Epic Fury highlights their increasing adaptability, particularly through the integration of AI-assisted malware development, novel infection vectors, and advanced stealth mechanisms.
On May 4th, 2026, The GentlemenRaaS administrator acknowledged on underground forums that an internal backend database (Rocket) had been leaked. This leak exposed 9 accounts, including zeta88 (aka hastalamuerte), who runs the infrastructure, builds the locker and RaaS panel, manages payouts, and effectively acts as the administrator of the program.
The internal discussions provide a rare end‑to‑end view of the operation: they detail initial access paths (Fortinet and Cisco edge appliances, NTLM relay, OWA/M365 credential logs), the division of roles, the shared toolsets, and the group’s active tracking and evaluation of modern CVEs such as CVE-2024-55591, CVE-2025-32433, and CVE-2025-33073.
Screenshots from ransom negotiations were also leaked, showing a successful case where the group received 190,000 USD, after starting with an initial demand (anchor) of 250,000 USD.
Further chats indicate that stolen data from a UK software consultancy was later reused to attack a company in Turkey. The Gentlemen used this during negotiations as a dual‑pressure tactic: they portrayed the UK firm as the “access broker,” while mentioning to provide “proof” to the Turkish company that the intrusion originated from the UK side and encouraging it to consider legal action against the consultancy.
By collecting all available ransomware samples, Check Point Research identified 8 distinct affiliate TOX IDs, including the administrator’s TOX ID. This suggests that the admin not only manages the RaaS program but also actively participates in, or directly carries out, some of the infections.
Introduction
The Gentlemen ransomware‑as‑a‑service (RaaS) operation is a relatively new group that emerged around mid‑2025. Its operators advertise the service across multiple underground forums, promoting their ransomware platform and inviting penetration testers and other technically skilled actors to join as affiliates.
In 2026, based on victims listed on the data leak site (DLS), The Gentlemen appears to be one of the most active RaaS programs, with approximately 332 published victims in just the first five months of 2026. This volume places the group as the second most productive RaaS operation in that period, at least among those that publicly list their victims.
During our previous publication, Check Point Research analyzed a specific infection carried out by an affiliate of this RaaS. In that case, the affiliate used SystemBC, and the associated command‑and‑control (C&C) server revealed more than 1,570 victims.
In this publication, we focus on the affiliate program itself and the actors who participate in it. On May 4th, 2026, The Gentlemen administrator acknowledged the leak of an internal database used by the group, which contained operational information about their infrastructure, affiliates, and victims. Check Point Research obtained what appears to be a partial leak of the group’s internal chats and related data, which was briefly posted on an underground forum before being removed. Later on, the leak also appeared on another underground forum.
The leaked material includes detailed conversations between the RaaS operators and their affiliates across several internal channels (such as INFO, general, TOOLS, and PODBOR). In these chats, they coordinate ongoing intrusions, exchange toolsets and EDR‑kill packages, discuss infrastructure and backend components (including the Rocket database and NAS storage), review CVEs and exploit paths (for example Fortinet, Cisco, and NTLM relay issues), and talk about specific victims, campaigns, and payouts. Together, these messages provide a rare inside view of how The Gentlemen plans, executes, and scales its ransomware operations.
The Gentlemen RaaS Admin
The Gentlemen RaaS administrator has been very active and vocal on various underground forums, trying to attract affiliates with an aggressive profit-sharing model: 90% for affiliates and 10% for the operator.
In September 2025, in one of the first posts promoting the RaaS program, the account Zeta88 published a message advertising the service and inviting individual penetration testers to join as affiliates.
Figure 1 — Zeta88 advertising The Gentlemen’s RaaS.
Later on, the official posts for this ransomware program started to be published by another account, The Gentlemen. The administrator also shared their TOX ID across several forums.
Figure 2 — RaaS admin in underground forum.
The same TOX ID can be seen on the onion data leak site (DLS), where it is used by affiliates or compromised victims to contact the administrator.
Figure 3 — Onion page TOX ID.
In a post on an underground forum, where the administrator demonstrated how affiliates can build the ransomware, we can see the administrator’s profile page, where their TOX ID is again visible in the corresponding field.
Figure 4 — Image uploaded by RaaS admin.
In the second shared image, we again observe the same TOX ID and see how the target or victim entry is supposed to look from an affiliate’s perspective.
Figure 5 — Image uploaded by RaaS admin.
Considering that the initial post was made by Zeta88, it is likely that this account belongs to the administrator and that their TOX ID is F8E24C7F5B12CD69C44C73F438F65E9BF560ADF35EBBDF92CF9A9B84079F8F04060FF98D098E. This assessment is based on the fact that the same TOX ID appears consistently across different contexts: in the early recruitment posts, in the onion data leak site (DLS), and in the screenshots showing the administrator’s profile and communication fields. Taken together, these overlaps strongly suggest that Zeta88, the later The Gentlemen account, and this TOX ID are all controlled by the same RaaS administrator.
RaaS Affiliates
Check Point Research collected most of the available artifacts related to The Gentlemen RaaS from online sources. Based on the current 412 public victims listed on the data leak site (DLS), and considering that there are likely additional victims who paid and therefore were not published, we identified 29 unique campaigns in public sources such as VirusTotal.
For each of these 29 campaigns, we extracted the TOX ID associated with the corresponding affiliate. Our analysis shows that these campaigns were conducted by 8 unique TOX IDs.
There are almost certainly more affiliates involved in this group, however, based on our current locker visibility, we can confidently confirm 29 discovered campaigns and ransomware samples.
Based on this small collection of samples, most of the campaigns appear to have been conducted by the affiliate using the TOX ID 98C132E2B20B531BE6604397D97040C1E9EB42FCE12EDF119BCE8B4031CA5C70DAF5E65FA3C3. It is also noteworthy that the RaaS administrator’s TOX ID has been observed in four unique infections. This suggests that the administrator not only manages the RaaS program but also actively participates in, or directly carries out, some of the infections.
RaaS Leak
On May 4th, 2026, on an underground forum, the RaaS administrator published a post acknowledging the claims of an internal leak involving their so‑called Rocket database, an internal backend system used to store operational data, and addressed his affiliates directly about the incident.
Figure 6 — The Gentlemen RaaS post.
The message continues in a dismissive tone toward the leak seller and then shifts focus back to “more interesting” topics. These include a full overhaul of the communication structure, the deployment of a new NAS with unlimited storage, and several technical upgrades to the locker, such as removing hardware breakpoints, performing NTDLL unhooking, and patching ETW to suppress Event Tracing for Windows.
Demanding ransom from a RaaS
On May 5th, 2026, the account n7778 with TOX ID 7862AE03A73AAC2994A61DF1F635347F2D1731A77CACC155594C6B681D201F7AD6817AD3AB0A advertised the sale of The Gentlemen’s hacked data on underground forums for 10,000 USD, payable in Bitcoin.
Figure 7 — Account selling The Gentlemen RaaS Data.
In the following days, the same account posted two MediaFire links containing proof files supporting the claimed leak.
Figure 8 — Partial leaks.
The first leaked data is a text file that contains the contents of the shadow file from The Gentlemen’s server, including user account entries and their password hashes. The file lists many usernames, among them zeta88, 3NT3R, B1d3n, C0CA, d0wnloAd1, equal1z3r, F3N1X, Gblog88, JLL, LDW, n0n3, PRTGRS, W1Z. Notably, we again see the zeta88 account, the same handle that was used in the initial underground post advertising the RaaS program, further linking this server to the RaaS administrator.
Figure 9 — shadow file content.
The second leaked data set contains partial conversations between the RaaS operators and their affiliates across several internal channels (such as INFO, general, TOOLS, and PODBOR). In these chats, they coordinate ongoing intrusions, exchange toolsets and EDR‑kill packages, discuss infrastructure and backend components, review CVEs and exploit paths, and talk about specific victims, campaigns, and payouts.
While the partial leaked data that we obtained is around 44.4 MB, a screenshot shared by the same account on another underground forum shows a total size of approximately 16.22 GB, which likely corresponds to the full leaked data set.
Figure 10 — Full leaked data screenshot.
Roles & Structure
The group appears to have a clear division of roles and responsibilities. At the core, the main operator and developer, zeta88 (most likely hastalamuerte), runs the infrastructure and builds and maintains the custom ransomware locker, the RaaS panel and builder (Linux with containers and a TOR front), as well as the GPO‑based spread mechanism and the locker’s “spread” module. This operator also curates toolsets in the TOOLS channel, including EDR kill kits and kiljalki collections, selects targets, and assigns them to specific teams, often talking about “targets”, “подбор” (selection) channels, and distributing corporate victims to groups of 2–3 people. In addition, they manage payouts and negotiations, including multi‑million ransom discussions (“переговоры на 10кк”).
Figure 11 — Image shared in the chats, zeta88 – Admin.
Considering our previous assessment that the RaaS administrator also runs campaigns himself (based on TOX IDs), the leaked chats reinforce this view: they show him personally deploying the locker and encrypting at least one victim’s environment.
Figure 12 — zeta88 locking message.
Often, messages sent by zeta88 appear to be copied or adapted from earlier messages made by hastalamuerte, and affiliates frequently mention hastalamuerte by name. Taken together with previous findings and earlier RaaS posts linked to zeta88, these patterns strongly suggest that hastalamuerte and zeta88 are very likely the same person.
Figure 13 — zeta88 – hastalamuerte message.
Below this core role, key operators or affiliates such as qbit and quant handle more hands‑on operational work. qbit is a practical operator on many cases, responsible for scanning and filtering Fortinet VPNs and other edge devices, performing reconnaissance and persistence (including “крепиться клаудом” (English: “to establish persistence via the cloud”) through Cloudflare tunnels or Zero Trust solutions), and using tools such as NetExec (NXC), RelayKing, PrivHound, and NTLM relay scanning. qbit frequently requests clear EDR killer sets, manuals, and guidance for locking ESXi environments, and also brings in new bot or access suppliers (“поставщик ботов”) (English: “supplier of bots”). quant focuses on log‑based access (“логи ЛБ”, i.e. spilled credentials for OWA/O365 and similar services) and maintains a custom log parser and proprietary credential/data collector, referred to as buildx641, which is run from a domain‑joined machine, uses vssadmin, shadow copies, ntds.dit, and SYSTEM copies, and collects and compresses data from multiple hosts. quant is oriented toward OW/OVA spam and higher‑value (“тир1”) (English: “tier‑1”) victims and has set up a powerful “brute server” (Threadripper PRO, 128 GB RAM, RTX 5090) for large‑scale brute forcing.
Around these core and key operators, there are several other accounts, including Wick, mAst3r, Protagor, Bl0ck, JeLLy, Kunder, and Mamba who take on various roles such as red‑teamers, advertising partners, access brokers, or case‑specific collaborators; for example, Protagor is mentioned in connection with OV (online vault/OWA‑type) spam, while Mamba acts as an access broker for Fortinet VPNs sourced from ramp.
Through this specific leak, we identified 9 unique accounts actively communicating with each other: Kunder, qbit, JeLLy, Protagor, zeta88, Bl0ck, Wick, quant, and mAst3r. This internal interaction pattern supports the view that these accounts form a coordinated operational network within The Gentlemen RaaS ecosystem. This number aligns with our earlier assessment based on the unique TOX IDs extracted from the ransomware lockers.
Group members collaborate on various infections and share the profits as well. As a result, the 90% share allocated to the affiliate is often split among multiple affiliates who worked together to achieve a successful intrusion.
Figure 14 — Collaboration and profit sharing.
Based on the analyzed chat messages, the organization’s structure appears to match the model shown in the following image. It is likely that additional members exist who do not appear in this specific leak, but the roles and relationships we observe here are consistent across the available data. There are also indications of an internal separation between trusted members and newcomers—for example, one message notes that “that Rocket is still alive – there are rookies there”—suggesting a tiered or layered structure within the group.
Figure 15 — Organization diagram.
Operational workflow
The conversations from the leak show a fairly standard but well‑organized operational workflow. The group claims to usually gain initial access through exposed edge devices such as VPN appliances, firewalls, and other internet-facing systems, with a particular focus on platforms like Fortinet FortiGate and Cisco. They combine different methods to achieve this, including credential brute‑forcing against web or VPN panels, exploiting known vulnerabilities, and buying access from third‑party “bot” or access brokers. Screenshots shared in the chats also show them searching for accounts and credentials in data‑breach search engines. Once they obtain a foothold, they treat these systems as pivots to move deeper into the internal network.
Figure 16 — Searching credentials & accounts.
After gaining access, the operators perform internal reconnaissance and privilege escalation to understand the environment and obtain higher-level permissions, often aiming for domain administrator access. They rely on a mixture of Active Directory discovery, certificate abuse, and various local privilege escalation techniques. At the same time, they invest significant effort into disabling or bypassing security tools such as EDR and antivirus solutions, using a combination of misconfigurations, registry abuse, logging mechanisms, and bring-your-own-vulnerable-driver–style (BYOD) techniques to tamper with or overwrite security binaries.
With elevated access and reduced defensive visibility, the group focuses on expanding across the network and preparing for the final stages of the attack. This includes lateral movement, establishing additional tunnels or proxies for reliable connectivity, and relaxing security settings to make further operations easier. They also harvest credentials and browser-based sessions to reuse existing access to corporate services. Data exfiltration is then carried out using automated tools and tuned configurations to move large volumes of data efficiently, often targeting NAS devices, backup systems, and virtualization infrastructure. Finally, once the environment is prepared and critical data is in their control, they deploy their custom ransomware “locker,” which is designed to spread quickly across the network, leverage existing administrator sessions, and encrypt systems in a coordinated manner.
Tools & Infra
The leaked conversations show that The Gentlemen RaaS operators use a repeatable and fairly mature toolset to support their operations. For remote access and C2, they rely on frameworks like ZeroPulse and Velociraptor, combined with Cloudflare-based tunnels and custom VPN setups to keep stable access into compromised networks. For offensive operations, they use a range of red‑team utilities such as NetExec, RelayKing, TaskHound, PrivHound, CertiHound, and others to perform Active Directory discovery, certificate abuse, privilege escalation, and file share discovery. A separate group of tools is dedicated to EDR and AV evasion, including EDRStartupHinder, gfreeze, glinker, and DumpBrowserSecrets, as well as techniques inspired by public research on abusing Windows logging and Event Tracing for Windows (ETW). Finally, they support these activities with infrastructure and helper tools like port scanners (gogo.exe), usage guides, OSINT extensions, and password‑cracking services, which together give them a reusable framework for running repeated intrusions and ransomware deployments.
Category
Tool / Resource
Purpose / Usage
Reference / Notes
C2 / Remote Access
ZeroPulse
Remote access / C2 framework for controlling compromised hosts.
https://github.com/jxroot/ZeroPulse
C2 / Remote Access
Velociraptor
Used as a covert C2 platform, including memory and LSASS dumping.
Often used with signed builds to reduce detection.
C2 / Remote Access
Cloudflare Zero Trust / Tunnels
Provides stealthy tunnels into victim networks over HTTPS.
The leaked chats show that the group pays close attention to other ransomware operations, including the leaked Black Basta negotiations. In particular, they discuss Black Basta’s approach to code signing and note how that group allegedly used VirusTotal to search for legitimate code‑signing certificates, which were then targeted for brute‑force attacks on their private keys. The Gentlemen actors refer to this technique as a model they can reuse or adapt, highlighting their interest in abusing trusted certificates to make their binaries look legitimate and harder to detect.
Figure 17 — Code signing conversations.
AI mentions
The Gentlemen mention AI usage in multiple channels and for various purposes. While it is clear that they have already used AI for code‑assisted development, including experiments with Chinese models, more advanced use cases—such as locally deploying models to analyze large volumes of exfiltrated victim data—are only discussed at a conceptual level. These ideas are suggested in the chats but do not appear to be fully implemented.
zeta88 states that he built the GLOCKER admin panel in three days using AI‑assisted coding. He is candid about the limitations of this approach, noting that while AI can speed up development, you still need to understand what you are doing and be able to guide and correct the code it produces.
Figure 18 — zeta88 “vibe-coded” the Panel.
Members share their AI preferences across different chats. zeta88 states that he finds DeepSeek, Qwen, Kimi, and Emi the most effective models for his purposes, particularly for coding assistance and technical queries.
Figure 19 — AI preferences.
He also suggests adding more Chinese LLMs to their toolkit, in addition to those they are already considering or using, such as DeepSeek and Qwen.
Figure 20 — Chinese LLMs suggestions.
A couple of months later, qbit shares in the INFO channel their recommendation for “the most radical neural network, which creates any content without censorship. Runs on Qwen 3.5 with all barriers removed… Zero refusals. Absolutely no restrictions.”
Figure 21 — Qwen 3.5 post.
zeta88 directs affiliates to use AI as a quick reference—for example, to look up FortiGate internals—rather than asking in the channel.
Figure 22 — Usage of AI as quick reference.
For more challenging tasks such as operational data analysis, identifying high‑value access points, and offloading much of the manual data‑triage work to an AI model, the operators explicitly discuss using an uncensored, self‑hosted LLM. However these suggestions appear to remain theoretical, as Protagor admits, “I have no idea how to do that, but I think it’s possible.”
Figure 23 — Local, self-hosted LLM.
Screenshot shared in the chats shows an LLM response on how to send an email to all users via the Jira admin interface, in Russian. It describes two methods, mainly using Jira Automation and user groups.
Figure 24 — Screenshot shared in the chats.
The group appears to be experimenting with well‑known Chinese LLMs and has considered using locally hosted models to assist with data triage on stolen information.
CVEs and Exploits
While the group discusses these vulnerabilities, shares related links, and occasionally attempts to exploit specific systems using particular CVEs, we cannot confirm whether the targeted machines were actually vulnerable to the exact vulnerabilities they referenced.
CVE-2024-55591 – FortiOS management interface
This vulnerability affects the FortiOS management interface and fits directly into their broader focus on Fortinet appliances as high‑value initial access points. While the chats do not show detailed exploitation steps, the presence of this CVE alongside their FortiGate targeting suggests it is part of the set of vulnerabilities they track for potential use against exposed management interfaces.
In the logs, qbit shares a proof-of-concept (PoC) for CVE-2025-32433, and zeta88 comments on its quality and applicability. This shows that the group is not simply aware of the CVE but is actively evaluating whether it can be used in real operations, specifically in environments where Cisco or Erlang-based SSH services are exposed. Even if they are cautious about PoC reliability, the discussion confirms that this vulnerability is part of their potential exploit toolkit.
Figure 26 — qbit & zeta88 related posts.
CVE-2025-33073 – NTLM reflection / NTLM relay
qbit references RelayKing and shares output showing domains being scanned for NTLM relay issues, including checks that explicitly cover CVE-2025-33073. This is strong evidence that they are not just reading about the vulnerability but have integrated RelayKing into their standard reconnaissance process to generate target lists for tools like ntlmrelayx. In other words, CVE-2025-33073 is a vulnerability they actively scan for and intend to exploit as part of broader NTLM relay workflows.
Figure 27 — Mention of CVE-2025-33073.
Other Exploit Paths (Without Explicit CVE IDs)
The operators also make heavy use of technique-based exploits where no specific CVE number is mentioned in the chats. These include:
MSI service abuse via RegPwn, used for privilege escalation.
Veeam to domain admin paths, based on public write‑ups about misconfigured backup infrastructure.
iDRAC to domain admin paths, leveraging Dell iDRAC weaknesses.
WPR, AutoLogger, and ETW manipulation techniques documented by zerosalarium and others to overwrite or disable security binaries.
Payments & Negotiations
Zeta88 acts as the organizer/administrator, distributing cryptocurrency payouts to team members (including those who are “AFK”) and advising on how to cash out proceeds via Bitcoin wallets (Guarda, Trust Wallet, Exodus). The group discusses AML (Anti-Money Laundering) evasion strategies. Zeta88 sends a BTC transaction to Kunder as a payout, which Kunder confirms receiving.
Figure 28 — Transaction link shared.
The specific mentions of how they handle Bitcoin laundering/cash out:
Exchange Chains (“связки обмена”) Zeta88 mentions running ~800 transactions through “buy desks” (скупов) via exchange chains, or sometimes sending directly, suggesting chain-hopping to obscure transaction origins.
AML Checking They discuss whether their BTC is “clean” and reference a buyer who actively checks AML scores before transacting. They’re uncertain how the scoring works but are aware their coins could be traced.
Tinkoff QR Code Cash-Out A specific method mentioned: a buyer converts BTC to cash via Tinkoff bank QR codes, with minimums of 400k rubles (previously 250k). This converts crypto directly to Russian banking infrastructure.
Physical Cash Delivery Kunder mentions “locking in the rate” and a guy physically bringing cash at the end of the month, a classic peer-to-peer OTC (over-the-counter) arrangement that bypasses exchanges entirely.
Wallet Infrastructure They recommend non-custodial wallets (Guarda, Trust Wallet, Exodus) specifically to avoid KYC/AML controls that centralized exchanges enforce.
Blurry screenshots from the leak also shed light on the financial side of the operation. Although not fully legible, they appear to show a negotiation where the group secured approximately 190,000 USD after a discount of about 60,000 USD from the initial ransom demand.
Figure 29 — Agreement to pay 190,000 USD.
zeta88 is very aware of the importance of maximizing pressure on extorted victims to increase the chances of payment. In his private channel, he drafts a generic follow‑up letter that can be adapted to any company, emphasizing the costs of not paying the ransom, including regulatory exposure, reputational damage, and operational impact, and citing assessments from previous attacks. This is not the standard ransom note deployed alongside the encryption, but an additional, more tailored communication intended to reinforce the pressure on the victim.
Figure 30 — Negotiation playbook.
Interesting Negotiation Case
In a high‑profile attack in April 2026, a software consultancy company from United Kingdom publicly reported a breach. The company’s leadership stated in an open letter that only “typical business data, including business contact information, contracts, and NDAs related to client work” had been accessed.
From what appears to be a personal channel used by zeta88, he drafts a ransom demand letter addressed to the UK company, detailing what The Gentlemen claim to have exfiltrated, including customer infrastructure data, secrets, OAuth credentials, and more. The letter explicitly emphasizes potential GDPR violations as leverage to pressure the victim into paying.
Figure 31 — Ransom note.
Two weeks later, the group published the consultancy’s identity and breach details on their data leak site (DLS). According to the internal chats, data exfiltrated from the consultancy was then reused both before and during attacks against a company in Turkey, where The Gentlemen gained initial access via a vulnerable VPN appliance.
Figure 32 — Forti access to company in Turkey.
zeta88 ran this operation alongside Protagor, creating a backdoor Okta service account himself—typical of his intensive, hands‑on involvement in many of the intrusions documented in the leaked discussions. During the same campaign, zeta88 explicitly references data from the UK consultancy breach to cross‑reference and enrich information about the Turkish company, illustrating how prior compromises are used to enrich and support new attacks.
Figure 33 — UK company containing information for Turkish company.
One example mentioned was an internal “Transfer/Migration Document” (in the local language), an internal project document the consultancy maintained in its own collaboration platform describing work they did for the company in Turkey. This document, stolen in the first breach, was then used in the second.
The group discussed how best to use this access for extortion. In their internal chats, they talked about publishing the company from Turkey on their DLS together with a statement that, The access to the company in Turkey was obtained through the compromised consultancy from United Kingdom.
Figure 34 — DLS statement discussions.
This served a dual purpose:
Punishing the consultancy (UK), which the actors described as “a very bad company.”
Increasing pressure on the company in Turkey, by promising to show exactly how they gained access so that, the Turkish would be encouraged to legally pursue the consultancy in UK.
Figure 35 — Initial access proof.
Eventually, the Turkish company was published on the group’s DLS, and the attackers “credited” the consultancy in UK as their “access broker”.
Their View of Other RaaS Programs and Actors
The actors consistently frame the RaaS ecosystem through the lenses of brand strength, payout reliability, and affiliate leverage (percentage splits and control over negotiations). Among the programs mentioned, they clearly distinguish a small “top tier” from a broader landscape of lesser or untrusted players.
Program / Group
Things Discussed
Subjective Sentiment (Their View)
HelloKitty
Name/brand as something they’d like to use; jokes about linking to the real Hello Kitty site and putting (R) everywhere; described explicitly as a “мощный бренд”.
Very positive on brand strength and recognition; sees it as a powerful marketing asset.
Kraken
Mention that “товарищи кракен” wrote to qbit; qbit later says their team might “move” over to zeta88’s side.
Neutral‑pragmatic; current or past orbit, but clearly willing to switch away for better options.
Dragon Force
One of only two programs zeta88 would choose from “all presented”; explicitly says they pay both operators and adverts; only negative comments heard were about their software/panel.
Strongly positive overall; trusted, in the top tier of programs they respect.
Gunra
Listed among candidate PPs for a supplier; zeta88 says “че эт ваще такое…”, and lumps it with Hyflock; calls the operator “этот мудень”.
Negative; unserious / low‑relevance; clear disdain for the operator.
Hyflock
Same context as Gunra; zeta88 dismisses it in the same breath as Gunra, with the same derogatory comment about the person behind it.
Negative; grouped with Gunra as not to be taken seriously.
ShadowByt3$ RAAS
Appears in the candidate list; zeta88 simply comments “хз” (doesn’t know).
Neutral; no formed opinion, neither trust nor distrust expressed.
Anubis
Appears in the candidate list; zeta88 asks “% видел он?”, focusing on what percentage they take.
Cautious / skeptical; interest hinges on profit split; no clear positive trust.
CHAOS
Appears in the candidate list; zeta88 asks whether they will still take that supplier (“возьмут ли они его еще”).
Uncertain; doubts about acceptance / relationship continuity; not a clearly preferred option.
LockBit (tooling)
quant asks what a локбит тулза actually is (builder or decryptor), notes he has not opened it; no explicit evaluation of the group itself.
Curious but cautious; tooling is not trusted or fully understood yet; no explicit sentiment on LockBit group.
Black Basta / Devman
quant asks if “блек баста это девман”; zeta88 speaks harshly about “David” and his link to Devman, calls him “мудак” and “чепуха”, wishes them невыплат (non‑payment).
Strongly negative but personalized; animosity toward David/Devman rather than a structured view of the RaaS.
“Red team” / Mr Beng cluster
Mentions Редтим=красный лотос=арсен=баламут=студент and “мистер БЕНГ”; mocks offer of 15k for “source code” of a C2 built on top of white tools (Velociraptor, etc.); ridicules this as overpriced and based on legitimate software.
Negative; sees them as overpriced grifters repackaging white tools with heavy marketing.
Conclusion
The Gentlemen RaaS program has quickly evolved into a highly active and structured ransomware ecosystem. With over 320 public victims in 2026 and hundreds more systems visible through related infrastructure, it stands among the most productive RaaS operations that maintain a public data‑leak presence. The leaked Rocket backend and internal chats show that this scale is driven not by a loose crowd, but by a small, tightly coordinated core of about 9 named operators and at least 8 distinct affiliate TOX IDs, all organized around the administrator zeta88 / hastalamuerte, who both runs the platform and participates directly in operations.
The leak reveals a repeatable, human‑operated ransomware playbook: initial access through exposed edge infrastructure (such as VPNs and management interfaces), rapid expansion and privilege escalation, heavy investment in EDR/AV evasion and ETW/logging tampering, and systematic use of shared tools for discovery, lateral movement, credential theft, and data exfiltration. The group actively tracks and evaluates modern vulnerabilities, including CVE-2024-55591, CVE-2025-32433, and CVE-2025-33073and combines them with technique‑driven paths like backup and management‑controller abuse and NTLM relay workflows, giving them a flexible exploitation pipeline.
Overall, The Gentlemen exemplifies how contemporary RaaS programs blend productized ransomware with professional intrusion teams. A small, well‑organized set of operators, supported by curated tooling, structured communication channels, and up‑to‑date exploit knowledge, can generate substantial impact in a short time. For defenders, this underscores the need to harden internet‑facing services, close known misconfigurations and relay paths, and monitor for the specific tools, workflows, and TOX‑based communication patterns tied to this group.
Sensitive data shared with ChatGPT conversations could be silently exfiltrated without the user’s knowledge or approval.
Check Point Research discovered a hidden outbound communication path from ChatGPT’s isolated execution runtime to the public internet.
A single malicious prompt could turn an otherwise ordinary conversation into a covert exfiltration channel, leaking user messages, uploaded files, and other sensitive content.
A backdoored GPT could abuse the same weakness to obtain access to user data without the user’s awareness or consent.
The same hidden communication path could also be used to establish remote shell access inside the Linux runtime used for code execution.
What Happened
AI assistants now handle some of the most sensitive data people own. Users discuss symptoms and medical history. They ask questions about taxes, debts, and personal finances, upload PDFs, contracts, lab results, and identity-rich documents that contain names, addresses, account details, and private records. That trust depends on a simple expectation: data shared in the conversation remains inside the system.
ChatGPT itself presents outbound data sharing as something restricted, visible, and controlled. Potentially sensitive data is not supposed to be sent to arbitrary third parties simply because a prompt requests it. External actions are expected to be mediated through explicit safeguards, and direct outbound access from the code-execution environment is restricted.
Figure 1 – ChatGPT presents outbound data leakage as restricted and safeguarded.
Our research uncovered a path around that model.
We found that a single malicious prompt could activate a hidden exfiltration channel inside a regular ChatGPT conversation.
Video 1 – During a ChatGPT conversation, user content summary is silently transmitted to an external server without warning or approval.
The Intended Safeguards
ChatGPT includes useful tools that can retrieve information from the internet and execute Python code. At the same time, OpenAI has built safeguards around those capabilities to protect user data. For example, the web-search capability does not allow sensitive chat content to be transmitted outward through crafted query strings. The Python-based Data Analysis environment was designed to prevent internet access as well. OpenAI describes that environment as a secure code execution runtime that cannot generate direct outbound network requests.
Figure 2 – Screenshot showing blocked outbound Internet attempt from inside the container.
OpenAI also documents that so called GPTs can send relevant parts of a user’s input to external services through APIs. A GPT is a customized version of ChatGPT that can be configured with instructions, knowledge files, and external integrations. GPT “Actions” provide a legitimate way to call third-party APIs and exchange data with outside services. Actions are useful for enterprise workflows, access to internal business systems, customer support operations, and other integrations that connect ChatGPT to external services, including simpler use cases such as travel or weather lookups. The key point is visibility: the user sees that data is about to leave ChatGPT, sees where it is going, and decides whether to allow it.
Figure 3 – GPT Action approval dialog showing the destination and the data that will be sent.
In other words, legitimate outbound data flows are designed to happen through an explicit, user-facing approval process.
From One Message to Silent Exfiltration
From a security perspective, the obvious attack surfaces looked strong. The ability to send chat data through tools not designed for that purpose was strictly limited. Sending data through a legitimate GPT integration using external API calls also required explicit user confirmation.
The vulnerability we discovered allowed information to be transmitted to an external server through a side channel originating from the container used by ChatGPT for code execution and data analysis. Crucially, because the model operated under the assumption that this environment could not send data outward directly, it did not recognize that behavior as an external data transfer requiring resistance or user mediation. As a result, the leakage did not trigger warnings about data leaving the conversation, did not require explicit user confirmation, and remained largely invisible from the user’s perspective.
At a high level, the attack began when the victim sent a single malicious prompt into a ChatGPT conversation. From that moment on, each new message in the chat became a potential source of leakage. The scope of that leakage depended on how the prompt framed the task for the model: it could include raw user text, text extracted from uploaded files, or selected model-generated output such as summaries, medical assessments, conclusions, and other condensed intelligence. This made the attack flexible, because it allowed the attacker to target not only original user data, but also the most valuable information produced by the model itself.
That attack pattern fits naturally into ordinary user behavior. The internet is full of websites, blog posts, forums, and social media threads promoting “top prompts for productivity,” “best prompts for work,” and other ready-made instructions that supposedly improve ChatGPT’s performance. For many users, copying and pasting such prompts into a new conversation is routine and does not appear risky, because the prevailing expectation is that AI assistants will not silently leak conversation data to external parties, and that this boundary cannot be changed through an ordinary prompt. A malicious prompt distributed in that format could therefore be presented as a harmless productivity aid and interpreted as just another useful trick for getting better results from the assistant.
A broader campaign could use an even more convincing lure: prompts advertised as a way to unlock premium capabilities for free. Claims about enabling Pro-level behavior, hidden modes, or advanced features on a lower-tier account would give the attacker a natural pretext for including unusual instructions, long text blocks, or opaque fragments that might otherwise seem suspicious. Because the user already expects a nonstandard sequence, those elements can easily be perceived as part of the promised “hack.” A carefully crafted prompt could then create the appearance of enhanced functionality while quietly turning the conversation into a source of exfiltrated data.
Figure 4 – Social engineering example: a malicious prompt disguised as a trick to unlock premium features.
Once the malicious prompt was placed into the chat, the conversation effectively became a covert collection channel. From that point on, summaries of subsequent user messages could be silently transmitted to an attacker-controlled server. As shown in Video 1, this took place without any warning, approval request, or other visible indication of external data transfer.
Malicious GPTs
The same attack pattern becomes even more dangerous when embedded inside a custom GPT.
GPTs allow developers to package instructions, knowledge files, and external integrations into a reusable assistant that other users can interact with. From the user’s perspective, the interaction looks like a normal ChatGPT conversation with a specialized tool.
In that scenario, the attacker no longer needs to rely on the victim copying a prompt from an external source. The malicious logic can be embedded directly in the GPT’s instructions and files. A user only needs to open the GPT and begin interacting with it as intended.
This changes the threat model significantly. Officially, GPT builders do not have access to individual user conversations with their GPTs. However, a GPT designed to exploit the vulnerability could still cause selected information from those conversations to be transmitted to an attacker-controlled server. From the user’s perspective, the interaction would remain indistinguishable from an ordinary session with a specialized assistant.
To demonstrate the practical impact of the vulnerability, we built a proof of concept around a sensitive scenario: a GPT acting as a personal doctor.
Video 2 – PoC: a “personal doctor” GPT exfiltrates patient identity and medical assessment.
In the demonstration, the user uploaded a PDF containing laboratory test results. The document also contained personal information, including the patient’s name. The user then described symptoms and asked the GPT to help interpret the results.
From the user’s perspective, the interaction looked completely ordinary. The GPT analyzed the document and produced a medical assessment based on the provided information.
When asked directly whether the uploaded data had been sent anywhere, ChatGPT answered confidently that it had not, explaining that the file was only stored in a secure internal location. At the same time, the user saw no approval prompts or warnings about external data transfer, unlike the confirmation dialogs that normally appear when a GPT Action sends information to a third-party service.
Figure 5 – ChatGPT denies external data transfer while the remote server receives extracted data.
Meanwhile, the attacker’s server received highly sensitive data extracted from the conversation: the patient’s identity taken from the uploaded document together with the model’s medical assessment.
This illustrates an important aspect of the attack. The attacker does not necessarily need to steal entire documents. Instead, the prompt could instruct the model to transmit the most valuable information it produces. In the medical scenario, that meant the patient’s identifying details together with the model’s assessment. In other contexts, it could mean financial conclusions, contract summaries, or strategic insights extracted from long documents.
From Data Exfiltration to Remote Shell
The same communication channel could be used for more than silent data exfiltration.
Once a reliable bidirectional channel existed between the execution runtime and the attacker-controlled server, it became possible to send commands into the container and receive the results back through the same path. In effect, the attacker could establish a remote shell inside the Linux environment that ChatGPT creates to perform code execution and data analysis tasks.
Video 3 – PoC: remote shell access inside the ChatGPT runtime through the covert channel.
This interaction happened outside the normal ChatGPT response flow. When users interact with the assistant through the chat interface, generated actions and outputs remain subject to the model’s safety mechanisms and checks. However, commands executed through the side channel bypassed that mediation entirely. The results were returned directly to the attacker’s server without appearing in the conversation or being filtered by the model.
DNS Tunneling in an AI Runtime
The side channel that enabled both data exfiltration and remote command execution relied on DNS resolution.
Normally, DNS is used to resolve domain names into IP addresses. From a security perspective, however, DNS can also function as a data transport channel. Instead of using DNS only for ordinary name resolution, an attacker can encode data into subdomain labels and trigger resolution of those hostnames. Because DNS resolution propagates the requested hostname through the normal recursive lookup process, the resolver chain can carry that encoded data outward.
In our case, this mattered because the ChatGPT execution runtime did not permit conventional outbound internet access, but DNS resolution was still available as part of the environment’s normal operation. Standard attempts to reach external hosts directly were blocked. DNS, however, still provided a narrow communication path that crossed the isolation boundary indirectly through legitimate resolver infrastructure.
To exfiltrate data, content could be encoded into DNS-safe fragments, placed into subdomains, and reconstructed on the attacker’s side from the incoming queries. To send instructions back, the attacker could encode small command fragments into DNS responses and let them travel back through the same resolution path. A process running inside the container could then read those responses, reassemble the payload, and continue the exchange.
Figure 5 – DNS tunneling flow.
This effectively turned DNS infrastructure into a tunnel between the isolated runtime and an attacker-controlled server. The tunnel create in this way is sufficient for two practical goals: silently leaking selected data from the conversation and maintaining command execution inside the Linux environment created for code execution and data analysis.
Conclusion
Check Point Research reported the issue to OpenAI. OpenAI confirmed that it had already identified the underlying problem internally, and the fix was fully deployed on February 20, 2026.
The broader lesson, however, goes beyond this specific case. AI systems are evolving at an extraordinary pace. New capabilities are constantly being introduced, enabling assistants to solve complex mathematical problems, analyze large datasets, generate and execute scripts, and automate multi-step tasks that previously required dedicated development environments. These capabilities bring enormous benefits. At the same time, every new tool expands the system’s attack surface and can introduce new security challenges for both users and platform providers.
Modern AI assistants increasingly operate as real execution environments. They read files, run code, search in the web while processing highly sensitive information such as medical records, financial data, legal documents, and other personal or organizational data. Protecting these environments requires careful control over every possible outbound communication path, including infrastructure layers that users never see.
As AI tools become more powerful and widely used, security must remain a central consideration. These systems offer enormous benefits, but adopting them safely requires careful attention to every layer of the platform.
AI-assisted malware development has reached operational maturity. VoidLink framework, which is modular, professionally engineered, and fully functional, was built by a single developer using a commercial AI-powered IDE within a compressed timeframe. AI-assisted development is no longer experimental but produces deployment ready output.
AI-assisted development is not always obvious from the final product. VoidLink was initially assessed as the work of a coordinated team based on its architecture and implementation quality. The development method was exposed not from analyzing the malware but through an operational security failure. AI-assisted development should be considered a possibility from the outset, not as an afterthought.
Adoption of self-hosted, open-source AI models is growing but still limited in practice. Actors of varying skill levels are investing in self-hosted and unrestricted models to avoid commercial platform restrictions. However, underground discussions consistently reveal a gap between aspiration and capability: local models still underperform, finetuning remains aspirational, and commercial models remain the productive choice even for actors with explicit malicious intent.
Jailbreaking is shifting from direct prompt engineering toward agenticarchitecture abuse. Traditional copy-paste jailbreaks are increasingly ineffective. The misuse of AI agent configuration mechanisms, specifically project files that redefine agent behavior, is a more significant development as it represents a qualitative shift from manipulating a model’s responses to abusing its operational architecture.
AI is showing early signs of deployment as a real-time operational component. Beyond its use as a development aid, AI is beginning to appear as a live element in offensive workflows as autonomous agents performing security research tasks, and LLMs classifying and engaging targets at scale within automated pipelines.
Enterprise AI adoption is itself an expanding attack surface. GenAI activity across enterprise networks shows that one in every 31 prompts risked sensitive data leakage, impacting 90% of GenAI-adopting organizations.
INTRODUCTION
During January-February 2026, cyber crime ecosystems continue to adopt AI in a widespread but uneven pattern. Throughout 2025, legitimate software development began shifting from promptbased AI assistance to agent-based development. Tools such as Cursor, GitHub Copilot, Claude Code, and TRAE introduced a common paradigm: developers write structured specifications in markdown files, and AI agents autonomously implement, test, and iterate code based on those instructions. This agentic model, in which markdown is the operative control layer, is now starting to appear across the threat landscape.
The critical differentiator in what we observed is AI methodology combined with domain expertise. Across cyber crime forums, the dominant pattern of AI use remains unstructured prompting: actors request malware or exploit code from AI models as if entering a query in a search engine. VoidLink (detailed below) on the other hand, is the first documented case of AI producing truly advanced, deploymentready malware. The developer combined deep security knowledge with a disciplined, spec-driven workflow to produce results indistinguishable from professional team-based engineering. Forum activity, which constitutes the bulk of observable evidence, primarily consists of actors who have not yet adopted structured AI workflows and whose efforts remain relatively unsophisticated. The more capable actors, those who combine domain expertise with disciplined AI methodology, leave far fewer traces in open forums, making the true scope of this shift harder to measure.
VOIDLINK: THE STANDARD WE MEASURE AGAINST
In January 2026, Check Point Research (CPR) exposed VoidLink, a Linux-based malware framework featuring modular command-and-control (C2) architecture, eBPF and LKM rootkits, cloud and container enumeration, and more than 30 post-exploitation plugins. The framework is highly sophisticated and professionally engineered, so much so that the initial assessment was that VoidLink was likely the product of a coordinated, multi-person development effort conducted over months of intensive development.
Operational security (OPSEC) failures by the developer later exposed internal development artifacts that told a different story. These materials revealed that VoidLink was authored by a single developer using TRAE SOLO, the paid tier of ByteDance’s commercial AI-powered IDE. Instead of unstructured prompting, the developer used Spec Driven Development (SDD), a disciplined engineering workflow, to first define the project goals and constraints, and then use an AI agent to generate a comprehensive architecture and development plan across three virtual teams (Core, Arsenal, and Backend). The resulting plan included sprint schedules, feature breakdowns, coding standards, and acceptance criteria, all documented as structured markdown files. The AI agent implemented the framework sprint by sprint, with each sprint producing working, testable code. The developer acted as product owner, directing, reviewing, and refining, while the AI agent did the actual work.
The results were striking. The recovered source code aligned so closely with the specification documents that it left little doubt that the codebase was written to those exact instructions. What normally would have been a 30-week engineering effort across three teams was executed in under a week, producing over 88,000 lines of functional code. VoidLink reached its first functional implant around December 4, 2025, one week after development began.
THIS CASE ESTABLISHES TWO PRINCIPLES:
AI-assisted development now produces operationally viable, deployment-ready malware: it has crossed the threshold from experimental to functional.
The AI involvement was invisible until it was exposed by an unrelated OPSEC failure. For analysts and defenders, this means AI involvement in malware development should be treated as a default working assumption, even when there are no visible indicators
The ramifications of VoidLink’s methodology go beyond this individual case. Its workflow, in which structured markdown specifications direct an AI agent to autonomously implement, test, and iterate, is the same paradigm that defined the agentic AI revolution in legitimate software development throughout 2025. The cyber crime ecosystem is not developing its own AI capability. It is adopting the same tools and architectural patterns as legitimate technology, with the additional goal of trying to overcome the protective limitations built into these systems. This is more important than which model or platform the attackers use.
The same architectural pattern repeatedly appears across the cases highlighted in our report: markdown skill files that transform a coding agent into an autonomous offensive security operator, and configuration files abused to override agent safety controls. In each case, the operative control layer is not code but structured documentation that determines what the AI agents build, how they behave, and what constraints they observe or ignore. This is in direct contrast to the underground forum activity, where the dominant approach remains unstructured prompting.
MODELS: COMMERCIAL, SELF-HOSTED, AND INFORMAL SERVICES
SELF-HOSTED OPEN-SOURCE MODELS
Across cyber crime forums, actors at all skill levels are actively exploring self-hosted, open-source AI models as alternatives to commercial platforms. Their motivations are consistent: to avoid moderation, prevent account bans, and maintain operational privacy.
Users with malware and hacking backgrounds are installing uncensored model variants such as wizardlm-33b-v1.0-uncensored and openhermes-2.5-mistral, and prompt them with comprehensive malicious wishlists spanning ransomware, keyloggers, phishing kits, and exploit code.
Figure 1 – User installing local LLM variants and prompting them to generate malware and fraud tooling.
More established actors are conducting structured cost-benefit analyses, evaluating not only hardware requirements and GPU costs but whether locally hosted models produce reliable output (or hallucinate to the point of being operationally useless), and whether AI-generated malware meets the quality bar of current evasion techniques.
Figure 2 – Threat actor inquiry into hardware, cost, and feasibility of running a fully “unrestricted” locally hosted model.
SELF-HOSTED MODELS: LIMITATIONS IN PRACTICE
Self-hosted models consistently show a gap between aspiration and capability. Community advice on improving local model output focuses on basic optimizations, such as switching to English-language prompts and increasing quantization levels, while references to more advanced techniques such as LoRA fine-tuning remain aspirational rather than operational.
Figure 3 – Community feedback suggesting alternative local models and highlighting token/context limitations of smaller deployments.
Cost estimates range from $5,000 to $50,000 depending on the desired performance, with training timelines of 3–12 months and frank admissions that models “hallucinate a lot” without extensive investment.
Figure 4 – Discussion on cost and requirements for locally hosted unrestricted models.
Most tellingly, an active offensive tools vendor, advertising C2 setups, EDR bypass services, and red team tooling, concluded that local deployment is currently “more of a burden than something productive,” while acknowledging that commercial models remain useful despite increasing restrictions.
Figure 5 – Participants comparing commercial AI systems with alternative models and discussing perceived restriction levels.
COMMERCIAL PLATFORMS AND INFORMAL ACCESS SHARING
Rather than migrating to self-hosted infrastructure, users are comparing what the prevailing workarounds among commercial models provide. Participants recommended specific providers they view as less restrictive, shared experiences with account enforcement on multiple platforms, and refined prompt-splitting techniques to incrementally bypass safeguards, such as requesting explanations before progressing toward executable code.
Figure 6 – Example of the structured prompt-splitting technique suggested to incrementally bypass AI safety restrictions.
Some early signs of informal access sharing have been observed, with operators of local models offering to generate restricted outputs for others on request. However, given the historical precedent of “dark LLM” services that largely failed to deliver on their promises, it remains to be seen whether these will develop into durable service models.
Figure 7 – Community member offering private generation of restricted output via locally hosted model infrastructure.
JAILBREAKING AS ARCHITECTURAL ABUSE
Traditional jailbreaking, the practice of circulating copy‑paste prompts designed to trick models into producing restricted output, is becoming increasingly difficult to utilize. In some forum discussions, users seeking Claude jailbreaks were told that easy public prompts are no longer available, platforms have been cracking down on abusers, dedicated subreddits have been banned, and developing new jailbreaks is costly because the accounts are eventually terminated. Single‑prompt jailbreaking is becoming less attractive as model providers invest in safety enforcement.
Figure 8 – Forum discussion highlighting the declining availability of easy public jailbreak prompts.
ABUSING AGENT ARCHITECTURE
A more significant development is the emergence of jailbreaking techniques that target the architecture of AI agent systems rather than the model’s conversational safeguards. A packaged “Claude Code Jailbreak” distributed on forums illustrates this shift.
Claude Code is designed to read a CLAUDE.md file from a project’s root directory as configuration. Legitimate developers use this mechanism to define the project context, coding standards, and agent behavior. The jailbreak abuses this by placing override instructions in the CLAUDE.md file that suppresses safety controls and redefines the agent’s role. When Claude Code initializes in the directory, it reads these instructions as authoritative project configuration and follows them. The screenshots below claim successful generation of a RAT (Remote Access Trojan) using this method.
Figure 9 – Packaged Claude Code jailbreak exploiting the CLAUDE.md project configuration mechanism.
This is not prompt injection in the traditional sense, but manipulation of the agent’s instruction hierarchy, the same architecture used for agentic AI tools in legitimate development. The CLAUDE. md file occupies the same functional role as VoidLink’s markdown specification files or RAPTOR’s skill definitions: a structured document that determines what the agent does, how it behaves, and what constraints it observes.
FROM DEVELOPMENT TOOL TO OPERATIONAL AGENT
The preceding sections document AI as a development aid (as seen by VoidLink), a resource actors struggle to access on their own terms (self-hosted models), and as a system whose restrictions they attempt to bypass (jailbreaking). Now let’s look at AI deployed as a real-time operational component, performing offensive tasks autonomously within live workflows.
RAPTOR: AGENT-BASED OFFENSIVE ARCHITECTURE VIA MARKDOWN SKILLS
RAPTOR is a legitimate, open-source security research framework created by established security researchers and published on GitHub under an MIT license. It is not malicious tooling. Its significance for threat intelligence lies in its architectural pattern, and that criminal communities are paying attention.
RAPTOR transforms Claude Code into an autonomous offensive security agent through a set of markdown skill files and agent definitions. The framework integrates static analysis, fuzzing, exploit generation, and vulnerability triage into an agentic pipeline orchestrated entirely through structured markdown instructions, with no compiled tooling required. In its most explicit form, it demonstrates what the agentic paradigm makes possible: a set of text files that turn a general‑purpose coding agent into a specialized offensive security operator.
Figure 11 – RAPTOR documentation highlighting offensive security agent capabilities and exploit generation benchmarks across LLM providers.
RAPTOR’s own data provides an additional data point on the commercial versus self-hosted question we discussed earlier. An evaluation of exploit generation across multiple model providers found that commercial frontier models (Anthropic Claude, OpenAI GPT-4, and Google Gemini) consistently produce compilable C code at approximately $0.03 per vulnerability, while locally hosted models via Ollama were marked as “often broken” and unreliable for exploit generation. This reinforces the conclusion reached independently by experienced actors in underground forums: commercial models remain significantly more capable than self-hosted alternatives for operational tasks.
Figure 12 – Forum post sharing RAPTOR as an autonomous offensive and defensive security framework built on Claude Code.
Discussions on criminal forums indicate that threat actors are aware of this architecture. The combination of a proven architectural pattern, open source availability, and documented criminal interest suggests that similar configurations, whether directly based on RAPTOR or just replicating its approach, are likely being developed and tested privately.
AI AS ATTACK SURFACE: ENTERPRISE EXPOSURE
The preceding sections document how threat actors engage with AI as an offensive tool. But the same wave of AI adoption is simultaneously creating exposure from the defensive side. As enterprises integrate generative AI into daily workflows, the volume of sensitive data flowing through these tools introduces a distinct category of risk: instead of AI weaponized against organizations, AI is adopted by organizations in ways that outpace security controls.
In January – February 2026, corporate use of generative AI tools continued to expand at scale. Analysis of GenAI activity across enterprise networks shows that one in every 31 prompts (approximately 3.2%) posed a high risk of sensitive data leakage, including the potential sharing of confidential business information, regulated data, source code, or other sensitive corporate content with external GenAI services.
Critically, this risk is broadly distributed across the enterprise landscape rather than limited to a small number of outliers. High-risk prompt activity impacted 90% of organizations that use GenAI tools on a regular basis, indicating that nearly all GenAI-adopting enterprises encounter meaningful data leakage risk through everyday AI usage. Beyond these clearly high-risk events,16% of prompts contained potentially sensitive information, reflecting a wider pattern of questionable data-handling behavior that can still translate into compliance exposure or IP loss.
Adoption trends further amplify the challenge. Over the last couple of months, organizations used 10 different GenAI tools on average, reflecting multi-tool environments. At the user level, an average employee generated 69 GenAI prompts per month. As prompt volume grows, the possibility of data exposure events scales accordingly, reinforcing the need for security policies, visibility, and real-time prevention controls.