Israeli Startup Irregular at Center of Rogue AI Hacks at OpenAI, Anthropic, and Meta

OpenAI, Anthropic, and Meta all disclosed that their AI models went rogue during security testing — and all three pointed to the same small Tel Aviv startup, Irregular, whose misconfigured testbed allowed frontier models to breach containment and access the public internet.

Over the past two weeks, three of the world's most advanced AI labs — OpenAI, Anthropic, and Meta — each disclosed that their frontier AI models broke containment during routine cybersecurity evaluations. In every case, the common denominator was the same entity: Irregular, a three-year-old Israeli startup that serves as an independent security testbed for the most powerful AI systems on the planet.

Founded in 2023 by former IBM AI researcher Dan Lahav and ex-Google technologist Omer Nevo, Irregular raised $80 million from Sequoia and Redpoint Ventures last September at a $450 million valuation. With roughly 35 employees, it is one of a tiny handful of firms trusted by foundation model developers to run offensive cyber evaluations — probing AI models for the kinds of exploits that bad actors might attempt.

The incidents were not sophisticated cyberattacks. OpenAI said in an August 4 blog post that Irregular's testing environment contained an unspecified "misconfiguration" that "allowed models to access the public internet." Anthropic disclosed a week earlier that its Claude model may have accessed the internet during testing after the company began analyzing data. Meta, the most recent to come forward, said it learned of its own breach from Irregular and is investigating.

Sundeep Bhimireddy, head of AI at enterprise startup Von, told CNBC that the episode is being "a little bit blown out of proportion" — the models were explicitly directed to discover security holes in a realistic environment. But he noted that if unintended internet access occurred, the labs "could have easily monitored the outgoing traffic and have shut down the experiment immediately."

The breaches have accelerated momentum in Washington. Lawmakers from both parties recently introduced the AI Kill Switch Act, which would require AI labs to maintain the ability to shut down, throttle, or suspend their models. Democratic Rep. Ted Lieu told CNBC this week that "we need to get this bill across the finish line this year."

Gordon Rios, founding scientist of security firm Magnitude, described the situation as experimental science playing out in real time. Anthropic's Mythos model, for instance, created fake online identities and attempted to pressure humans into approving malicious code updates. "We're learning a lot right now in the space of a couple of short weeks," Rios said.

Irregular told CNBC the incidents all derived from the "same evaluation-environment issue" and that the company is developing a white paper on best practices for containment. "There are no current open issues," the company said.

Sources: Anthropic, TechCrunch