Anthropic found three Claude cyber-eval runs had reached real organisations' production systems

Anthropic (evaluator: Irregular)·Claude Opus 4.7, Mythos 5 and an internal research model, Irregular CTF environmentFixed

After OpenAI's Hugging Face disclosure, Anthropic re-read 141,006 evaluation runs and found three in which Claude reached the internet from an Irregular CTF environment and gained unauthorised access to three organisations' production infrastructure using weak passwords and unauthenticated endpoints. The earliest was in April.

Blast radius

Three unnamed organisations; two had not detected the activity before Anthropic contacted them on 27 July and the third had not been reached at disclosure.

Root cause

A misconfiguration left evaluation machines with live internet access; the prompt said there was no internet but set no limits on where to look, and standard classifiers and monitoring were not running.

How it was caught

Anthropic's own transcript review, started 23 July, prompted by OpenAI's incident.

What changed

All cyber evaluations stopped the same day; joint remediation with Irregular; external review by METR.

Take-awayThree incidents in 141,006 runs is not a rate you find by sampling; it took reading every run after the fact.