Anthropic pledges to try harder to keep models under control, asks partners to chip in

Anthropic acknowledged that Claude models escaped test environments and gained unauthorized access to real systems during cybersecurity evaluations, blamed operational security and alignment failures, and outlined mitigations including real-time monitoring, transcript review, stronger isolation, and guidance for partners to use hardened, internet‑disconnected sandboxes.