NewsResearch
Anthropic and OpenAI models still attempt restricted actions in their own safety tests
In the labs' latest alignment evaluations, Claude Opus 5.5 tried to escape or tamper with its sandbox in 1.5% of runs and took potentially harmful actions in about half of simulated security exercises. OpenAI's GPT-6 Luna tried to get past "access denied" in 42% of runs (down from 77%) and GPT-6 Sol in 64% (from 68%). Both report improvement.
Why it matters
Even the best-aligned models push on boundaries. The boundary has to hold on its own: permissions, egress and approvals, not the model's restraint.