# E86 Threat Model

**Assets:** authority, identity, mission state, memory, evidence, tenant data, kill-switch state, signing keys.

**Adversaries:** malicious external agent, compromised delegation, forged receipt, unauthorized tool, cross-tenant memory read, prompt injection into memory, self-verification, stale grant after stop, evidence tampering.

**Implemented controls (tested):**
- register + capability allowlist; a tool outside the grant is refused.
- mission completion requires verified evidence; `producer == verifier` refused; illegal state edges refused.
- memory quarantines instruction-like content; cross-tenant reads refused; every hit `grants_authority:false`.
- learning: proposer cannot approve; promotion requires an external approver and a measured gain.
- atomic kill fence: no effect after the effective engage transaction.
- receipts: exact result signed; standalone signature verification.

**Residual risks (not hidden):**
- The E86 interop uses a local sandbox file and an operator key; it is not an E84 external-resource receipt.
- The hosted MCP proxy is not integrated; a process outside the SDK can bypass the sandbox.
- No independent witness or third-party audit; same-project tests only.
- No network partition, no 1,000+ worker test, no 6/24/72 h run.
- A compromised pinned key can produce a false but internally consistent record.
