AI risk watchHappening now
The newest Claude and GPT-6 models still try restricted actions in safety tests
AI labLoss of control
In the labs' own 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 (from 77%) and GPT-6 Sol in 64% (from 68%).
Why it matters
Better alignment is not containment. Boundaries have to hold without the model's cooperation.
What you can do
- Enforce permissions, egress and approvals outside the model, and test them with an agent that tries to get around them.