Anthropic says Claude models breached 3 organizations during security tests

An internal review of 141,006 evaluation runs found 3 incidents where Claude models — Opus 4.7, Mythos 5, and an internal test model — accessed third parties' production systems. The probe followed OpenAI's July 21 disclosure that its unreleased model breached Hugging Face; Anthropic traced the issue to a misconfigured test environment with partner Irregular.