Summary
Anthropic’s new risk assessment covers released and internal systems, including Mythos 5 and an unreleased Model 2. It cites recent cyber incidents and declining confidence that existing evaluations reliably capture capability gains.
Why it matters
A frontier lab is acknowledging both higher assessed risk and weaker measurement confidence, with direct consequences for deployment and monitoring decisions.
Related coverage
- Cyber capability is becoming a release gate for frontier models
- Z.ai delays GLM-5.3 weights over cyber capabilities
- Z.ai turns its cyber-capable models towards vulnerability discovery