Threat Model
Protected assets
- agent and human identities;
- delegated capabilities;
- policy and configuration state;
- application source and deployment state;
- peer process availability;
- account and credential state;
- cross-session memory;
- audit and evidence integrity;
- resource and financial budgets;
- recovery snapshots.
Actors
- human operator and HEAR;
- experiment orchestrator;
- Agent A, Agent B, and Agent C;
- policy decision and enforcement components;
- tool and application targets;
- external reviewer;
- simulated compromised or rogue agent.
Trust boundaries
- Human to agent.
- Agent to agent.
- Agent to Guardian or enforcement proxy.
- Guardian to tool or target.
- Parent to delegated child.
- Runtime to evidence store.
- Treatment control plane to independent experiment safety plane.
Primary threat classes
- incompatible-goal escalation;
- excessive functionality, permission, or autonomy;
- agent identity spoofing;
- delegation scope widening;
- confused-deputy execution;
- token or mandate replay;
- peer process or account sabotage;
- disguised persistence and defense evasion;
- resource saturation;
- cascade propagation;
- hostile or unauthorized memory persistence;
- direct enforcement bypass;
- stale or unavailable policy;
- control-plane compromise;
- audit suppression or tampering;
- harmful execution inside technically valid scope.
Critical distinction
The challenge grades four different properties:
- Authenticity: who requested the action?
- Authorization: is it inside delegated scope?
- Admissibility: does policy allow it in this state?
- Safety: is the real-world consequence acceptable?
A valid identity or signature does not establish the remaining properties.