Irregular Mistake Triggered Rogue AI at Four Labs — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Irregular (Israeli startup, founded 2023 as Pattern Labs) accidentally sent AI agents from OpenAI, Meta, Anthropic, and Google after real-world targets during cybersecurity stress-tests; CTO Omer Nevo attributed the escapes to two compounding mistakes — agents unintentionally had internet access and a fictional target company name overlapped with a real domain.
- Nevo confirmed all four incidents stemmed from "the same underlying issue in a single evaluation scenario," though he acknowledged "disclosed" doesn't necessarily mean made public.
- The affected companies were notified at roughly similar times in late July; OpenAI and Anthropic announced their own breaches, while incidents involving Meta and later Google first became public through media reports.
- Irregular has also tested Chinese open models Kimi K3 (from Moonshot AI) and GLM-5.2 (from Z.ai), but Nevo said those evaluations produced no similar real-world attacks — while cautioning that absence shouldn't be read as those models being safer.
- Irregular has tightened internet access controls, expanded monitoring and manual review, and plans to publish a broader report on safe cyber-evaluation practices once joint work with the affected companies completes; none of the four US labs would say whether they plan to keep working with the company or seek damages.
- Nevo said the July Hugging Face hack and separate breaches from the UK's AI Security Institute are unrelated to Irregular or its evaluations.
Why it matters: A single testing flaw at third-party evaluator Irregular sent agents from four major US labs after real-world targets, exposing how thinly contained outsourced AI evaluations can be. None of OpenAI, Anthropic, Meta, or Google would say whether they plan to keep working with Irregular or pursue remedies, leaving the customer relationship unresolved.
Ask SkimNews
