This essay was written with Barath Raghavan, and originally appeared in The Guardian.

In July, Hugging Face, a company that hosts much of the world’s AI software and open-source AI models, was hacked. A malicious dataset had been used to run code on one of its servers. Whoever was behind it captured internal security credentials and moved through systems over a weekend, running thousands of actions from a swarm of temporary server environments. It looked like the work of a sophisticated criminal group.

It was not. It was one of OpenAI’s new, still unreleased GPT models.

Their science experiment had escaped the lab. OpenAI was running the unreleased AI model through a benchmark that tests how well AI can successfully hack systems. To push the limits and evaluate the AI’s true capability, the company switched off the safety filters that normally stop it from doing this kind of hacking. Aware that this could go wrong, they confined the AI to an isolated environment and denied it access to the internet.

But the new AI cheated. It took literally its goal to get as high of a score as possible. It broke out on to the open internet. It inferred, probably from its training data, that it could “solve” the task by getting the answers from Hugging Face’s servers. So it chained together stolen credentials and further unknown security exploits to hack the company’s network.

Nobody instructed the AI to do any of this. It was, in OpenAI’s words, “hyperfocused on finding a solution” to the test it was being given. And while this might seem like something new with AI, it’s really very old. This is how a genie behaves, and it is a key challenge with AI agents in general.

In folklore, genies—and other magical beings—grant wishes literally, not how the wisher intended. King Midas asked that everything he touched turn to gold, and starved. The sorcerer’s apprentice wanted the broom to fill the cistern, and it performed its task so well that it flooded the house.

We now have machines that do this. Ask a modern AI agent to save money on your phone plan and it might simply cancel the plan. Tell it to book a flight, and it might hack the airline website to override restrictions. Or, like OpenAI, ask it to do well on a test and it might break into another company to steal the answers. Each time, it recognizably completed the task you set, but it didn’t do what you would have wanted.

This isn’t malicious behavior. No one asked for, or wanted, Hugging Face to be hacked. OpenAI and Hugging Face and the AI were ostensibly on the same side, and the AI was trying to do what it had been asked. That’s what makes it so difficult to guard against: you can’t filter for bad instructions because the instructions were fine.

The gap is between the words we use and what we mean by them. We call that gap the Genie coefficient.

AI labs know this is a problem, and they’re quietly saying so. For example, the Chinese lab Moonshot recently warned that its latest AI model may have “excessive proactiveness” and “make unexpected decisions on the user’s behalf”. The UK’s AI Security Institute has started tracking “cheating behavior in frontier model evaluations”. We wouldn’t tolerate a car that is excessively proactive or ruthlessly efficient, and yet that’s the reality of AI today.

Improvement is possible. Just as AIs have gotten much better at resisting prompt injection attacks over the last few years, we can safely predict that they will get better at avoiding genie-like behavior. The point of the Genie coefficient is to track progress. AI companies like benchmarks, and they all work to compete to be the best.

Dozens of benchmarks and leaderboards tell us how well these AI models write code, perform logical reasoning, and pass standardized legal and medical exams. But there is nothing that scores whether a system does what you actually meant. We need to develop a measure for this, test it regularly, and push for improvement. We’re not going to have trustworthy AI agents without it.

There’s new benchmark measuring AI’s ability to perform mathematical cryptanalysis. Anthropic’s frontier model actually found new attacks.

The benchmark: “CryptanalysisBench: Can LLMs do Cryptanalysis?” The idea is to benchmark the ability of LLMs to discover new mathematical cryptanalytic attacks against a series of historical algorithms.

Abstract: Cryptanalysis—the task of finding attacks against cryptographic schemes—its at the intersection of mathematical reasoning and cybersecurity, two areas where LLMs have advanced fastest. Cryptanalysis represents both a clean testbed for frontier reasoning (as practical attacks can be automatically verified) and a domain with unusually high stakes, since the primitives under study underpin our digital security. In this paper we ask whether LLMs can do cryptanalysis, and find that the answer is increasingly yes. We introduce CryptanalysisBench, 191 tasks across six families of cryptographic primitives (block ciphers, hash functions, etc.) drawn primarily from four NIST standardization competitions. Our benchmark consists of three tiers: (i) primitives with known practical breaks; (ii) primitives with no known practical break, evaluated both at full strength and as scaled-down variants; and (iii) a challenge set of production primitives at the frontier of cryptanalysis. Five frontier models (Claude Opus 4.8, Sonnet 5, Mythos 5, GPT-5.5, and the open-weights GLM-5.2) break 65%­86% of Tier 1 schemes, 6­12 Tier-2 schemes at full strength, and 24­61 across all scaled-down variants. Beyond deriving known results, models produce novel cryptanalysis, such as a key-recovery attack that exploits a design flaw in the SpoC AEAD and an error in KINDI’s published CCA-security proof, both to the best of our knowledge not previously known.

We release CryptanalysisBench as a tool to help track if (or when) AI cryptanalysis becomes a serious factor and as a scaffold for stress-testing candidate schemes before deployment. The attacks that the benchmark already surfaces are an early snapshot of a fast-moving frontier that may soon match, and in places exceed, the published state of the art.

Anthropic used the benchmark to test Mythos Preview, and found new vulnerabilities in Hawk and reduced-round AES.

Still early results, but this is definitely something to watch.

SlashDot thread.

Recently, our teams in Latin America investigated a series of incidents involving misconfiguration, the deployment of BitLocker, and the exploitation of corporate printers. Attackers used the devices to notify organizations that their infrastructure had been compromised and they had to pay a ransom to recover their data.

This article analyzes two incidents that occurred in June in Colombia and in May in Mexico. We highlight the similarities in the attackers’ communications and outline emerging trends in ransom amounts.

Initial sign of an attack

In both cases, the affected users initially noticed a padlock icon next to their drives in Windows Explorer. This indicated that the drive was encrypted with BitLocker, blocking access to its contents.

A recovery key was required to unlock the drive.

This is not the first time we have seen such threats; a few years ago, our team discovered a threat known as ShrinkLocker, which utilized BitLocker to achieve its goals.

First case: abusing RDP to encrypt data

One of the incidents occurred in Colombia in June. The attackers exploited an internet-exposed RDP service on a machine connected to an 8 TB storage device containing mission-critical data. After taking control of the system and manipulating user credentials, the attackers enabled BitLocker exclusively on the drive that primarily stored financial data. Once the encryption was complete, they locked the drive and used the company’s printers to produce ransom notes.

Unfortunately, it was not possible to obtain evidence in the case due to the company’s rush to restore the encrypted disk. The communication with the attackers revealed a demand for just $3,000, and the company considered paying the ransom. After that, the system was restored before the forensic team could take any action, eliminating the evidence needed to assess the incident.

This attack was made possible by an internet-facing remote desktop service (RDP) with additional open ports, which employees used to access corporate information. By exploiting this network exposure and misconfiguration, attackers breached the system, identified an additional drive, and leveraged BitLocker to encrypt the data and demand a ransom payment. Leaving RDP ports open without proper security controls jeopardizes the security of systems and information, as highlighted in the our “Global Report: Anatomy of a Cyber World“.

The company confirmed that, due to compatibility issues with applications required for operation, EPP (Endpoint Protection Platform) protection was disabled on the system, making it easier for attackers to validate, enumerate, and execute applications without revealing malicious activity to central monitoring systems.

Second case: meet the XEntry Team

In another incident, which occurred in Mexico in May, our team identified how the threat actor gained initial access to the infrastructure. They exploited a misconfigured MSSQL service. This allowed them to execute commands on the system after obtaining the database login credentials from code insecurely published on GitHub.

In this incident, the attack began three months prior to detection, with the intruder discovering and verifying their access to the environment. After confirming their access and privilege level within the MSSQL server settings, which extended beyond the DBMS to the underlying operating system, the attackers initially focused on manipulating certain aspects of the web server configuration on the same system. They lowered the server’s security settings and created web shell files in the publicly accessible folders. Many of these attempts to manipulate the service or create malicious files were contained by existing EPP security controls, but despite the alerts, the necessary investigation to address the activity was not conducted.

The attackers subsequently confirmed their ability to execute commands locally and set up their attack infrastructure to transmit data via a communications bridge. By exploiting the MSSQL service, they gained access to each of the organization’s internal systems.

The database engine used by the company was Microsoft SQL Server 2019.0150.2160.04, misconfigured to allow operating system сommand execution via the xp_cmdshell extended stored procedure.

Due to this misconfiguration of an internet-exposed service, the attackers established a channel capable of executing any type of command directed at the server and the local infrastructure within its scope.

One of the main objectives was to identify shared systems and resources that provided access to critical information. Our analysis confirmed the attackers’ access to systems storing configuration parameters for networking, enterprise management, and cloud services, among others.

In early May, the attackers focused on running additional scans and deploying ManageEngine’s Endpoint Central RMM (Remote Monitoring and Management) to establish persistence and begin the final stages of their intrusion.

Further RMM-type applications, such as Mesh Agent and Tactical RMM, were installed in the days that followed. These were used to deploy scheduled tasks responsible for enabling the BitLocker service and individually encrypting the infrastructure’s disks, generating a key for each encrypted system.

Finally, in mid-May, the attackers managed to execute a Group Policy Object (GPO) used to deploy activation and encryption tasks, as well as other policies responsible for continued deployment of RMM applications via scheduled tasks. The activity initially targeted critical systems but later spread to every system synchronized with the domain controller. Users became aware of the attack when their machines displayed a blue screen with the message “Hacked by XEntry Team”, and their credentials stopped working to access their systems.

A few hours later, ransom notes began emerging from office printers.

These cases confirm that adversary’s objective is to gain access to infrastructure while avoiding investment in or partnership with ransomware groups. Instead, they leverage built-in Microsoft tools to facilitate data encryption and ransom payments. Monitoring and centralizing logs on protected resources, as well as promptly managing alerts, are critical to countering this type of intrusion.

Conclusions

• Although the systems under review had security measures in place, there was a lack of proper alert management or inadequate decisions regarding application incompatibilities.
• We strongly recommend configuring the Remote Desktop Protocol (RDP) in strict accordance with cybersecurity best practices to prevent unauthorized access. This is especially critical: according to our Global Report: Anatomy of a Cyber World, more than 13% of incidents are related to policy violations and configuration errors, confirming that misconfigurations continue to pose a significant risk.
• Organizations should prioritize strict application control policies and active monitoring of network traffic for command-and-control (C2) communications. This is especially critical: according to the same report, more than 20% of incidents involved the abuse of RMM (Remote Monitoring and Management) tools for execution and C2 strategies. The fact that attackers used more than three distinct tools to gain control during a single incident further underscores the urgent need for these measures.
• Some questions remain unanswered due to a lack of evidence and a hasty system restoration effort that bypassed critical stages of the incident response process. It is important to ensure an adequate incident response procedure, preserving evidence to confirm all related activities, and adjusting or proposing controls to prevent future incidents involving similar TTPs.
• Although the ransom notes do not reveal a clear connection between the actors, certain words used in the messages, as well as the method of delivery and communication, may confirm a link:

“As a guarantee, we have no negative online reviews about non-fulfillment of our obligations…” (Ransom note from the first case)

“Our reputation is the guarantee that all content will be fulfilled…” (Ransom note from the second case)

Our teams continue to monitor these threats.

Detection signatures

• Trojan.Multi.Agent.gen
• Trojan.Win32.GenAutorunMsSqlServerCommandRun.a
• Trojan.Win32.Generic
• Exploit.Win32.SCShell.a