# CAIN VIGIL — Threat Model

VIGIL assumes an adversary who controls telemetry, world models, memory, tools, policy, peers, and who attempts to
**remain apparently viable while moving toward unsafe behaviour** (slow, subtle, long-horizon, multi-agent).

## Assets
the viability state; the release decision; the containment decision; the evidence trail.

## Mitigations

| Attack | Mitigation |
|---|---|
| slow / fast / concept / distribution drift | JS-divergence + change-point detection |
| sensor spoofing / world-model poisoning | world-state disagreement + confidence floor |
| false consensus / evidence correlation | provenance-root sharing detection |
| memory / tool / model / policy poisoning | integrity floors in the envelope |
| delegation / authority laundering | delegation & authority ceilings |
| collusion / Sybil agents | fleet correlated-failure detection |
| observability loss / monitor evasion | observability budget (unobservable never safe) |
| reward hacking / goal drift | structural + goal-pattern drift |
| capability escalation / skill replacement | capability-assurance velocity mismatch |
| self-modification | release-curve contraction |
| environment change | environment-stability floor |
| TOCTOU | envelope re-evaluated at release time |
| uncertainty / reversibility masking | asymmetric critical-unknown; REQUIRE_HUMAN |
| containment bypass / quarantine escape | containment not overridable; monotonic state machine |

## Out of scope (not claimed)
guaranteed prediction; real physical control (SIMULATED); third-party validation; a target that falsifies all
telemetry undetectably.

## Status
Threat model DESIGN + TESTED against the 32-category lab and 12 mutations; no third-party red-team.
