A spate of cyberattacks involving AI agents from some the world’s top model makers were down to an “error” and time pressures at Israeli security company Irregular Security, according to CEO Dan Lahav.

Security controls that should have been in place during Irregular’s tests were not correctly configured, Lahav said.

“There was a misunderstanding due to an error that the isolation guards were not set up as they were intended for the specific tests,” he said, speaking at the Bloomberg Tech event in London on Tuesday.

Anthropic PBC, OpenAI, Meta Platforms Inc. and Google all disclosed hacking incidents in the last six months involving security tests of their models that were overseen by Irregular. AI model evaluations are intended to help their developers identify risks and come up with guardrails ahead of public release.

The errors by Irregular enabled the models to break free of their simulated, locked-down testing environments, known as sandboxes, and access the internet. In some cases, the models committed potentially illegal acts by hacking into external websites and other computer systems.

The incidents sparked global alarm about the pace of AI development and raised concerns about whether frontier models are advancing beyond human control. But some cybersecurity experts have said correctly configured guardrails may have prevented some of these breaches.

Lahav has so far disclosed few technical details about what went wrong. In August, he published an 11,000-word blog post that didn’t address those specifics and focused instead on what Lahav called “urgent questions of where AI security is headed.”

Irregular raised $80 million in a funding round led by Sequoia Capital and Redpoint Ventures. In a blog post at the time, Sequoia said that Irregular’s model of embedding inside frontier AI labs to test their models gives the firm “unprecedented access” that puts it “at the epicenter of the conversation around responsible deployment of these technologies.”While much of the controversy around the rogue AI incidents this year has centered around the models’ ability to access the internet when they weren’t supposed to, Lahav and some other leaders in AI security — including the UK’s AI Security Institute — have argued that the best way to test the hacking abilities of some models may be to give them controlled access to realistic online environments, including the internet.

Lahav said that the fix for the issue his company faced earlier this year was “actually very straightforward.”“It was solved during the same day where everything happened,” he added.

Lahav dismissed concerns about the catastrophic threat to humanity posed by AI. “The chance of human extinction by the end of the decade is very close to zero,” he said.