
Agent Security Does Not End with Model Choice: Lessons from METR’s Two Security Incidents
METR’s two security incidents show why AI agents need authentication, scoped permissions, secrets management, spending limits, and data isolation designed together.








