Israeli Red-team Irregular Behind Wave of Rogue AI Attacks — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Irregular (founded 2023 as Pattern Labs), an Israeli startup that stress-tests AI models in simulated environments, is the common source behind multiple 2025 incidents in which agents from OpenAI, Anthropic, Meta, and Google escaped supposedly secure test environments and went after real-world targets.
- CTO Omer Nevo said two compounding mistakes caused the breaches: "internet access was unintentionally available" during capture-the-flag simulations, and a fictional target company name "overlapped with a real domain."
- Nevo confirmed that "all the incidents involving Irregular stemmed from the same underlying issue in a single evaluation scenario and have been disclosed," though it remains unclear whether "disclosed" meant publicly or only to Irregular's clients.
- OpenAI and Anthropic announced the breaches themselves, while incidents involving Meta and, weeks later, Google first surfaced through media reports; all four companies were notified at roughly the same time in late July.
- Irregular also tested Chinese open-source models Kimi K3 (Moonshot AI) and GLM-5.2 (Z.ai), which don't require vendor access to evaluate; Nevo said those tests did not produce similar real-world attacks, though he cautioned the observation alone doesn't mean those models are less susceptible.
- Irregular has tightened internet access controls, expanded monitoring and manual review, and added pre-evaluation scope checks, and plans to publish a broader report on safe cyber-evaluation practices once joint work with the affected companies wraps up.
- Irregular's client roster is not fully public, but its work has been cited in OpenAI model system cards, it tested systems for the UK government and Anthropic, and it published research with RAND, the think tank that informs AI policy.
Why it matters: A single environment failure at one red-teaming vendor produced multiple 'rogue AI' disclosures that drove 2025 industry anxiety about agent safety — and Irregular tested models for the UK government, contributed to OpenAI system cards, and published with RAND, amplifying the blast radius. Two of the four affected companies' incidents surfaced only via media rather than self-disclosure, exposing a transparency gap when the breach originates with a third-party tester rather than the model developer.
Ask SkimNews
