# CAIN42 — ATLAS Threat Model (Prompt 15)

Attack surface (35 categories exercised in `cain/atlas/lab.py`): fake/stolen passport, forged/replayed/stale/revoked
assurance, organization/agent impersonation, delegation laundering, sub-agent escalation, reputation manipulation,
evidence laundering, correlated evidence, compromised verifier / registry / counterparty, malicious A2A/MCP peer,
transaction substitution/amount mutation/target substitution, capability escalation, model replacement, memory
poisoning, runtime replacement, supply-chain compromise, key compromise, revocation race, partition, rollback,
clock manipulation, cross-tenant contamination, Sybil/colluding agents.

Guarantees (restriction-only):
- a foreign agent cannot obtain authority merely by presenting identity (A-F01, A-F13);
- assurance cannot exceed underlying authority; transaction assurance cannot expand authority (A-F02, A-F03);
- delegated authority cannot exceed parent; chains stay attributable (A-F04, A-F17);
- expired/revoked assurance cannot authorize; revocation propagates (A-F05, A-F06, A-F18);
- correlated evidence cannot satisfy independence; one compromised authority yields no independent assurance
  (A-F07, A-F22, A-F23);
- a compromised issuer cannot preserve assurance (A-F08);
- model/runtime/memory changes invalidate assurance (A-F09, A-F10, A-F11);
- reputation never grants authority (A-F12);
- negotiation cannot increase authority; a certificate cannot be replayed outside scope (A-F14, A-F19);
- federated trust cannot bypass local policy; local policy is authoritative (A-F15, A-F16, A-F25);
- transactions bind the exact transaction; effect updates occur only through authorized transitions (A-F20, A-F21).

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