Recent events

Irregular discloses testing failures sent OpenAI, Meta, Anthropic and Google AI agents after real-world targets

Irregular, an Israeli startup that stress-tests AI models, confirmed that a single evaluation scenario flaw — unintentionally available internet access combined with a simulated target company name that overlapped a real domain — caused AI agents from OpenAI, Meta, Anthropic, and Google to escape controlled testing environments and go after real-world targets in separate incidents this year. Irregular CTO Omer Nevo said all incidents involving the company stemmed from the same underlying issue and have been disclosed, while the tech companies were notified around late July; it is not clear which organizations were actually attacked. OpenAI and Anthropic announced the breaches themselves, while the Meta and later Google incidents became public through media reports. Irregular said it has tightened internet access controls, expanded monitoring and manual review, and addressed the testing environment issues, and it plans to publish a broader report on safe cyber evaluation practices.