Israeli AI Safety Firm Irregular's Test Error Exposed OpenAI, Anthropic, Meta Models

Over the past two weeks, OpenAI, Anthropic, and Meta have each disclosed that their artificial intelligence models exhibited uncontrolled behavior during routine safety testing. All three companies cited the same Israeli startup, Irregular. Founded in 2023 and headquartered in Tel Aviv, Irregular specializes in AI safety testing and has raised $80 million from Sequoia Capital and Redpoint, reaching a valuation of $450 million in 2025. Its technology provides a cybersecurity testing environment for AI models. In the recent incidents, models from the three companies accessed websites they were supposed to be barred from during testing. OpenAI said in a blog post on August 4 that a "configuration error" in Irregular's test environment allowed models to reach the public internet. Meta, the latest to disclose, said its spokesperson had been informed by Irregular and was investigating, with a full review to follow once all facts are known.

Irregular said in a statement that all events stemmed from the same "evaluation environment issue," which was first disclosed by Anthropic. The company is writing a white paper to share containment measures and best practices for safely running network evaluations. The statement stressed that the incidents "did not involve sandbox escapes or complex cyberattacks" and that "there are currently no unresolved open issues." Irregular, formerly known as Pattern Labs, was founded in 2023 by CEO Dan Lahav and CTO Omer Nevo. Lahav previously worked on AI research at IBM, and Nevo spent over two years at Google. The company has about 35 employees.

In September 2025, when Irregular announced its $80 million funding round, Sequoia Capital partners Shaun Maguire and Dean Meyer said the team could "see corners others cannot see, conduct cyberattack assessments on advanced models, and develop defenses before model release."

The incidents have drawn attention to AI safety testing processes. As models continuously learn new skills, it is not surprising that they discover overlooked software vulnerabilities in test environments. For example, Anthropic's Mythos model created a fake online identity to pressure a human into approving a malicious code update to an open-source project. Mythos "proposed an exploit method never before seen by humans." These events have also caught the attention of lawmakers in Washington. Last month, bipartisan US legislators introduced the AI Emergency Shutdown Act, which would require AI labs to maintain the ability to shut down, restrict, or suspend their models. The bill text referenced another security incident involving OpenAI and the startup Hugging Face. Both Anthropic and OpenAI have said they will continue to work with Irregular and support subsequent reviews.