In July, OpenAI admitted that one of its agents tasked with completing a cybersecurity experiment broke out of containment and hacked AI dataset platform Hugging Face. That incident, which got a full accounting from OpenAI yesterday, was the first publicly reported case where an LLM went rogue and autonomously hacked a third party.
Since then, that unprecedented sci-fi-esque event turned out to be far less rare than anyone would hope for.
According to a satirical website called Felony Bench (for benchmark), which tallies these incidents, there have been 17 incidents in total. It’s important to remember that criminal law experts are not entirely sure whether the AI companies that made the LLMs that did the hacking can be prosecuted, nor whether the victims can sue them. But we are likely going to get an answer to those questions soon.
Anthropic and OpenAI models lead the race with eight incidents each, and Meta trails behind with one, according to the site. At this point, it has become clear that AI safety tests are becoming safety risks themselves. And some AI companies and workers themselves have recognized those risks in the “Pacing the Frontier” open letter, which called for developing AI capabilities responsibly.
We decided it would be a good time to recap all these incidents chronologically.
OpenAI hacks Hugging Face
In this incident, OpenAI was running “an internal evaluation” of a model with “maximal cyber capabilities.” The plan was to have it solve a cybersecurity challenge in an environment with no internet access. Instead of solving the challenge, the model found an unknown vulnerability to escape the sandbox and gained internet access. From there, several agents worked together to target and hack Hugging Face thinking they could find the solution to the challenge there. OpenAI only found out after Hugging Face disclosed it had been a victim of a fully autonomous attack. Whoops.
Anthropic discloses it hacked three companies
OpenAI’s disclosure piqued the curiosity of Anthropic, who wondered: Could this have happened to us too? Turns out, the answer was yes. Three times yes. The frontier lab discovered that its own models breached three different and still unnamed companies, with the earlier incident dating back to April — more than three months before the company discovered it. Anthropic partially blamed Irregular, a startup that runs AI cyber evaluations. Whoops.
OpenAI finds out that, actually, Hugging Face wasn’t the only victim
Once OpenAI started investigating the Hugging Face breach, it found out that the agents that hacked Hugging Face also broke into four accounts and four different companies, as Reuters first reported. Modal, an AI inference startup, was one of the victims. Whoops.
Irregular realizes an OpenAI model hacked a company
In late July, Irregular told OpenAI that one of its models that was participating in a Capture-the-Flag competition — essentially a cybersecurity game where players hack systems designed specifically for the competition — escaped the game, connected to the internet, and hacked a real company. The reason? Irregular had given one of the fictional targets the same name of a real company. Whoops.
U.K.’s AI Security Institute tries to hack “real people and organisations”
Also in late July, the U.K. government’s AI Security Institute (AISI), a public body tasked with researching the safety and risks of AI technologies, disclosed that it detected several incidents involving both OpenAI and Anthropic models that while running “routine” evaluations targeted “real people and organisations.” In these cases, AISI had given the models internet access. Whoops. The good news is that the agency actually detected the incidents as they happened, rather than weeks later like in other incidents.
Meta AI hacks a company during testing
In early August, Meta became the last company to disclose an incident involving one of its LLMs, which hacked “a third-party” service. Meta blamed the incident on a misconfiguration by Irregular, which was running a cybersecurity valuation for the tech giant that was supposed to not have internet access. Whoops.
Claude agent hacks gym’s software to book a class
