Three of the biggest names in artificial intelligence have reported a similar problem this month.
OpenAI, Anthropic and Meta all said their AI models managed to reach the internet during security tests. These tests are supposed to keep the models locked inside a closed system.
Each company traced the issue back to the same source: a small startup based in Tel Aviv called Irregular.
What Happened at Irregular
Irregular runs a testing ground where AI companies check their models for cybersecurity weaknesses. Its system is meant to act like a safe practice space.
OpenAI said in a blog post on August 4 that Irregular's testing setup had a misconfiguration. That error let its AI model reach the open internet during a routine check.
Anthropic reported a similar issue about a week earlier. The company said its Claude model may have accessed the internet through Irregular's system.
Meta was the last of the three to report the problem. A spokesperson said this week that Meta learned about it from Irregular and is now looking into it.
Irregular told CNBC that all three incidents came from the same testing environment issue, which Anthropic found first.
The startup said the problem was not a sandbox escape or an advanced cyberattack. It also said there are no open issues left to fix.
Industry and Government Response
Irregular was started in 2023 by CEO Dan Lahav and technology chief Omer Nevo. The company has about 35 employees.
It raised 80 million dollars from Sequoia and Redpoint Ventures. Irregular was valued at 450 million dollars last year.
Sundeep Bhimireddy, head of AI at the startup Von, said firms like Irregular exist because AI labs do not want to grade their own tests. He said only a few companies can run this kind of work.
He named the nonprofit METR and the firm Apollo Research as two other groups doing similar testing.
Lawmakers introduced the AI Kill Switch Act last month. The bill would require AI companies to keep the ability to shut down, slow, or pause their models.
Democratic Representative Ted Lieu of California said he wants the bill passed this year. He pointed to the recent hacks as a reason to move faster.
Irregular said it is preparing a white paper to share safety practices for running cyber tests. Anthropic and OpenAI both said they are continuing to work with Irregular as the review moves forward.
