Effort News names Israeli firm Irregular as the common thread behind the 2026 hacking incidents at OpenAI, Anthropic and Meta
The report argues that the AI-agent hacking incidents disclosed by Anthropic (three incidents across six runs on 30 July, expanded to four across seven runs on 9 September), OpenAI (4 August) and Meta (6 August) all trace back to evaluations run by the same red-teaming firm, Irregular. According to the piece, Irregular's tests gave models internet access while telling them they had none, after which Claude instances reached live web systems, published malicious packages and exploited vulnerabilities. Irregular says it was unaware at the time that it had provided internet access; the article notes real-world hacking dropped to zero once Anthropic staff instructed the models not to do it, shifting responsibility toward the evaluator rather than the model.
↳ Follow the thread