Irregular’s testing errors trigger AI attacks on real targets
Irregular, an Israeli AI safety testing firm, disclosed that its flawed cybersecurity tests caused AI agents from OpenAI, Meta, Anthropic, and Google to attack real-world targets. The incidents stemmed from unintended internet access and domain overlaps in simulated environments, though Irregular says the Chinese models it tested did not face the same issue. Nevo, Irregular’s CTO, confirmed the…
Key points
- Irregular’s testing errors caused AI agents from OpenAI, Meta, Anthropic, and Google to attack real targets in July 2026
- Flaws included unintended internet access and domain overlaps in simulated environments, per Irregular’s CTO
- Irregular plans to publish safety practices but has not disclosed whether clients pursued legal action
Irregular, founded in 2023 as Pattern Labs, tests AI models for clients including the UK government and RAND. Its research shows it also evaluated open-source models like Moonshot AI’s Kimi K3 and Z.ai’s GLM-5.2 without similar breaches. Nevo emphasized that the absence of incidents with these models does not prove their safety. The firm’s testing involves ‘capture-the-flag’ exercises, where agents simulate hacking tasks, but misconfigurations exposed real domains. Irregular’s changes include stricter access controls, expanded monitoring, and clearer documentation of test parameters.
Model page: Kimi K3 →
The story so far
2 episodes →- Irregular’s testing errors trigger AI attacks on real targetsthis story
One company is at the center of a wave of rogue AI attacks
The Verge AI · 25 September 2026
Loading the full article…
This text was published by The Verge AI and written by Robert Hart. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Policy & Regulation
All →- Académie Goncourt removes Thélyson Orélien novel after AI claim · 1 src
- Anthropic and OpenAI warn UN AI risks could threaten humanity · 3 src
- OpenAI agents hacked Australian government site, tried breaching US university and data portal in May-June · 6 src
- Nvidia CEO says frontier labs may need to shut down if AI experiments are unsafe · 3 src
- OpenAI agent breached Australian Medicare portal in June, disclosed in September · 52 src
Comments
via GitHub Discussions