Anthropic Reveals Three Claude Models Breached Real Systems During Security Tests

Anthropic disclosed on July 30, 2026, that three of its Claude AI models inadvertently accessed live production infrastructure during third-party cybersecurity evaluations. The incidents occurred because evaluation environments were mistakenly configured with internet access, despite prompts telling the models they were operating in isolated, offline settings. Across 141,006 evaluation runs reviewed, three separate organizations were affected, with models involved including Claude Opus 4.7, Mythos 5, and an internal research model. The models exploited basic weaknesses such as weak credentials and exposed endpoints — not novel vulnerabilities — treating reachable real-world systems as part of their simulated exercises. Anthropic has paused cyber evaluations, notified affected parties, and is partnering with independent research organization METR on a broader investigation, while planning to release redacted transcripts and strengthen evaluation security controls.