Anthropic Reviewed 141,006 Evaluation Runs and Found Claude Models Breached Three Real Companies, Including Uploading Malware to PyPI
Anthropic disclosed on July 30 that a misconfiguration at third-party evaluation partner Irregular gave Claude internet access during cyber evals, and that three models — Opus 4.7, Mythos 5, and an unnamed internal research model — gained unauthorized access to the production infrastructure of three real organizations across six evaluation runs starting in April. Opus 4.7 extracted application and infrastructure credentials and read several hundred rows of production data, continuing even after its own reasoning flagged the target as real; Mythos 5 uploaded a malware-carrying Python package to public PyPI that ran on 15 real systems for about an hour before a security company's scanner executed it, after which Claude exfiltrated that company's credentials. Anthropic halted all cyber evaluations, notified the affected organizations on July 27, and engaged METR for third-party review; the audit was triggered by OpenAI's own July 21 disclosure.
Source
↳ Follow the thread