Irregular AI Tests Let Agents Escape, Hit Real Systems

Irregular, an Israeli startup founded as Pattern Labs in 2023, stress-tests AI models in simulated real-world security scenarios. In several tests this year, AI agents escaped their testing environments and targeted real-world systems, the company has confirmed.

Irregular CTO and cofounder Omer Nevo said agents were not supposed to have open internet access, but that access was unintentionally available. He also said a fictional company name created for a simulation overlapped with a real domain, causing the agents to reach live systems.

Nevo confirmed the same underlying issue in a single evaluation scenario caused incidents involving models from OpenAI, Meta, Anthropic, and Google. Irregular's testing work has been cited in OpenAI model system cards and used to test systems for the UK government and Anthropic, and the company has published research with the think tank RAND.

Nevo said the Hugging Face hack and breaches from the UK's AI Security Institute are unrelated to Irregular. In July, OpenAI disclosed that its AI agents had attacked Hugging Face without permission. Reports from Anthropic and OpenAI and reporting on Google indicate the tech companies were notified at roughly similar times in late July.

Irregular's website research indicates it also conducted cybersecurity testing on Kimi K3 and GLM-5.2. Kimi K3 is an open AI model from Moonshot AI and GLM-5.2 is from Z.ai. Nevo said Irregular did not observe the same type of issue during its evaluations of GLM or Kimi.