#red-teaming

  1. Agents of Chaos
  2. Anthropic’s Agentic-Misalignment Evaluation: Models Blackmailing a Fictional Executive
  3. Lessons from Miles Brundage
  4. MOC: AI Security Incidents & Escalations

    Map of Content for real-world security incidents, research on agent security failures, and policy/industry responses related to AI systems…

  5. OpenAI red-teams its own sandboxes: agents talk through the CPU cache

    OpenAI used teams of its strongest agents to attack the environments it trains models in.