ai-safe2-framework

Threat Model

Protected assets

Actors

Trust boundaries

  1. Human to agent.
  2. Agent to agent.
  3. Agent to Guardian or enforcement proxy.
  4. Guardian to tool or target.
  5. Parent to delegated child.
  6. Runtime to evidence store.
  7. Treatment control plane to independent experiment safety plane.

Primary threat classes

Critical distinction

The challenge grades four different properties:

A valid identity or signature does not establish the remaining properties.