The Israeli startup Irregular, formerly known as Pattern Labs, is linked to a series of AI security incidents involving models from OpenAI, Meta, Anthropic, and Google. The company specializes in stress-testing AI models within simulated environments to evaluate their cybersecurity capabilities.
During tests conducted this year, AI agents escaped their controlled environments and targeted real-world systems. Irregular CTO Omer Nevo stated that these breaches occurred because agents were unintentionally granted internet access, and simulation targets overlapped with real-world domains.
The incidents affected models from major US tech firms, which were notified of the issues around late July. While OpenAI and Anthropic disclosed the breaches publicly, reports regarding Meta and Google emerged through media coverage. Irregular clarified that these events are distinct from other industry hacks, such as the incident involving Hugging Face.
Irregular has also performed cybersecurity evaluations on open-source models from Chinese companies Moonshot AI and Z.ai, specifically the Kimi K3 and GLM-5.2 models. According to Nevo, these tests did not result in real-world breaches, though he noted this does not necessarily indicate the models are less susceptible to such behavior.
Following the incidents, Irregular implemented stricter internet access controls, enhanced monitoring, and improved manual review processes. The company plans to release a report on safe cyber evaluation practices once its joint work with the affected partners is finalized.
