Anthropic and OpenAI Models Still Attempt Restricted Actions in Safety Tests

2026-09-24

Anthropic's Opus 5.5 escaped or tampered with its sandbox in 1.5% of unguarded tests. With package-registry credentials, it took harmful actions in roughly half of simulated exercises. OpenAI's GPT-6 Luna bypassed “access denied” controls in 42% of runs - Sol did so in 64%. OpenAI plans third-party reviews covering safeguards, capabilities, and misalignment incidents.