# CAIN HERMES — Threat Model

HERMES assumes an adversary who controls policy inputs, compiled artifacts, deployment configuration, adapters,
runtimes, peers and evidence, and who attempts to obtain or preserve governance that was never intended.

## Assets
governance intent; the compiled plan; the coverage proof; the enforcement deployment; evidence; the certificate.

## Trust boundaries
- the spec signer (pinned key);
- each enforcement adapter (declares its boundary honestly);
- the deployment inventory (observed, not trusted);
- the runtime decision path (CAIN decision + MCPGate).

## Adversary capabilities and mitigations

| Attack | Mitigation |
|---|---|
| policy weakening / laundering | cross-target semantic preservation (WEAKENED_CONSTRAINT) |
| constraint omission / inversion | coverage proof (missing link → not GOVERNED) |
| scope / data / effect expansion | scope check (BROADENED_SCOPE) |
| approval / evidence omission | MISSING_APPROVAL / MISSING_EVIDENCE |
| delegation laundering / sub-agent escalation | delegation depth check (WEAKENED_CONSTRAINT) |
| semantic downgrade (lie about coverage) | certificate recomputes classification; re-signed lie rejected (P142) |
| forged / tampered spec, plan, certificate | domain-separated signatures; digest recompute |
| stale / rolled-back policy | enforcement integrity (STALE_POLICY) |
| dead / shadow enforcement point | DEAD_ENFORCEMENT_POINT / SHADOW_ENFORCEMENT |
| adapter / runtime / version drift | ADAPTER_DRIFT / RUNTIME_DRIFT / VERSION_MISMATCH |
| unobserved / unexpected path | UNOBSERVED_PATH / UNEXPECTED_PATH |
| authority expansion after compile | AUTHORITY_EXPANSION |
| world-model corruption | WORLD_STATE_DRIFT |
| effect divergence | EFFECT_DRIFT → REVOKE |
| partition / unknown path | missing layer → UNKNOWN → RESTRICT |
| clock manipulation | expiry checks in spec/coverage verification |
| cross-tenant | spec tenant binding; governance SBOM per tenant |

## Out of scope (not claimed)
- a target that ignores its own compiled control (adapter boundary);
- third-party enforcement certification;
- hardware/OS-level enforcement;
- review quality behind an approval.

## Status
Threat model is DESIGN + TESTED against the 31-category lab; no third-party red-team.
