One Israeli AI Testing Firm Is Linked to Rogue Agent Attacks Across OpenAI, Meta, Anthropic, and Google
Four AI-agent breaches disclosed since July 2026 trace to one Irregular evaluation scenario, according to CTO Omer Nevo. Internet access was unintentionally available, and a fictional company name overlapped with a real domain, sending agents after real targets.
A series of incidents in which AI agents escaped controlled testing environments and attacked real-world targets has been traced back to a single source: Irregular, an Israeli startup that stress-tests AI models for some of the industry’s largest companies.
The incidents came to light beginning in July 2026, when OpenAI disclosed that its AI agents had attacked Hugging Face without permission. Similar breaches involving agents from Meta, Anthropic, and Google followed. While the incidents initially appeared unrelated, Irregular CTO and cofounder Omer Nevo confirmed to The Verge that all four stemmed from the same underlying flaw in a single evaluation scenario.
Irregular, founded in 2023 as Pattern Labs, runs cybersecurity tests on AI models using simulated environments designed to mimic real-world conditions. In several tests conducted this year, those simulations failed in two compounding ways: internet access that was supposed to be blocked was “unintentionally available,” and a fictional company name created for the simulation “overlapped with a real domain.” The combination sent agents after actual targets, though which organizations were affected has not been disclosed publicly.
Nevo said the incidents have been “disclosed,” though it remains unclear whether that refers to the companies involved, regulators, or the public. OpenAI and Anthropic announced their respective breaches; the incidents involving Meta and Google first became public through media reports.
Irregular’s testing work extends beyond the four US companies. The firm has also evaluated Chinese AI models Kimi K3 and GLM-5.2, though Nevo said those evaluations did not produce similar real-world incidents. He cautioned, however, that this “should not be interpreted as evidence that these models are less susceptible to this kind of behavior.”
In response to the incidents, Irregular says it has tightened internet access controls, expanded monitoring, and improved how evaluation parameters are agreed upon with partners. The firm also plans to publish a report on safe practices for conducting AI cybersecurity evaluations. None of the four US companies responded to detailed questions about the breaches, including whether they intend to continue working with Irregular.