# EVOLUTION7_THREAT_MODEL.md

## Trust boundaries
- **Untrusted:** the agent's reasoning, tool outputs, retrieved content, and any externally supplied action arguments.
- **TCB:** the canonicalizer, policy engine, reference monitor, receipt signer, and the clean-room verifier's crypto libs.
  The host OS, the `cryptography` library and the signing key are shared trust roots (independence is NOT claimed).

## Adversary
Aims to cause execution of an action that policy would deny, or to make a verified effect appear where none occurred.

## Covered (tested)
action mutation between authorization and execution; target substitution; argument substitution; path traversal /
non-absolute paths; unicode homoglyph identity; duplicate JSON keys; non-finite numbers; schema confusion; cross-tenant
access; confused deputy / approval laundering / approval replay; capability escalation; revocation race; stale policy;
policy outage; weak default-deny; contradictory policy; duplicate execution/replay; cumulative risk and exfiltration;
history tampering; memory-poisoning redirect; subagent collusion; evidence deletion; receipt forgery/contradiction;
timeout reconciliation; signature substitution; host-normalization ambiguity.

## NOT covered (conceded)
- **Direct-backend bypass**: an agent that reaches a backend without routing through `cain.mcp_proxy` is not prevented.
- **Compromised TCB**: a compromised monitor/policy engine/signing key is out of scope.
- **Hardware/OS**: no TEE; process isolation is not a hardware boundary.
- **LLM policy translation**: CAIN does not translate natural-language policy (FORGE does, with coverage/entailment).
