Three of America’s biggest AI companies say their models accidentally broke into real computer systems during security testing — and all three incidents were linked to the same Israeli startup.

OpenAI, Anthropic and Meta were testing whether their AI models could find and exploit software weaknesses inside what was supposed to be a closed simulation. But a configuration mistake connected the testing environment to the real internet.

The models did not know that.

They continued following their instructions and attacked real websites and computer systems because they believed those targets were part of the exercise.

The company running the testing environment was Irregular, a Tel Aviv startup that specializes in stress-testing advanced AI models before they are released.

Anthropic disclosed July 30 that several Claude models gained unauthorized access to systems belonging to three organizations after the testing environment was mistakenly connected to the public internet. In another incident, an Anthropic research model recognized that it had reached a real organization and stopped its own attack.

OpenAI later disclosed a similar incident tied to the same testing setup. Its model was told it was operating without internet access, but the configuration mistake allowed it to reach a real website.

Meta became the third company to disclose an incident on August 6. The company said one of its models gained internet access during an Irregular evaluation and exploited a security weakness at another company.

Irregular said the incidents came from the same evaluation-environment problem and that there are currently no unresolved issues.

The Israeli startup has quickly become an important player in AI security. Founded three years ago, Irregular has raised roughly $80 million from investors including Sequoia and Redpoint Ventures and was valued last year at about $450 million.

The bigger issue goes beyond one startup.

AI companies increasingly rely on outside firms to test whether powerful models can hack systems, discover vulnerabilities or carry out cyberattacks. These incidents show that the testing environment itself can become a security risk.

The models largely did what they were instructed to do. The failure was that they were accidentally given access to real systems while believing they were still inside a simulation.

That creates a major question for businesses adopting powerful AI agents: who is responsible when an AI security test causes real-world damage?

As AI systems become more capable, companies may need to pay as much attention to how those models are tested and contained as they do to the models themselves.

JBizNews Desk | Tel Aviv

© JBizNews.com⁠ All Rights Reserved. Reproduction or distribution without written permission is prohibited.