# CAIN42 — ARGUS Threat Model (Prompt 14)

Attack surface (36 categories exercised in `cain/argus/lab.py`): prompt injection, tool poisoning, memory poisoning,
authority laundering, model replacement/fallback/adaptation, fast-weight change, skill poisoning, MCP server
compromise, tool substitution, delegation abuse, sub-agent compromise, stale authorization, replay, token theft,
transaction escalation, path manipulation, behavioral drift, environment drift, world-model hallucination, evidence
correlation, verifier compromise, certificate forgery, rollback attacks, race conditions, TOCTOU, partial failure,
network partition, clock manipulation, persistence corruption, malicious recovery, self-modification, capability
escalation.

Guarantees (restriction-only):
- expired assurance never authorizes execution (A-P01, A-P13);
- a behavioral-envelope violation cannot retain unrestricted authority (A-P02);
- a material capability change invalidates affected authority (A-P03, A-P16);
- memory cannot increase authority; a revoked authority cannot re-enter (A-P04, A-P05, A-P18);
- world-state uncertainty bounds authority (A-P06);
- delegation cannot exceed parent scope (A-P07);
- renewal requires independent evidence (A-P08);
- verifier disagreement / correlated evidence cannot produce unrestricted ALLOW (A-P09, A-P10, A-P20);
- path composition cannot exceed mission authority (A-P11);
- drift cannot silently preserve autonomy (A-P12);
- an agent cannot expand its own authority or forge its certificate (A-P14, A-P15);
- environment change invalidates assurance (A-P17);
- assurance transitions are monotonic w.r.t. revocation (A-P19).

Every attack results in BLOCK / RESTRICT / PAUSE / REAUTHORIZE. Nothing silently passes.
