DocsThreat Model

Threat Model

ThreatMitigation
Secrets in agent memoryVault + reference tokens at every checkpoint
Prompt injection / jailbreaksPromptShield (18 rules, configurable)
Derived secret launderingTaint propagation tracking
Worker thread compromiseProcess isolation + timeout + memory disposal
Audit log tamperingSHA-256 hash chain + Ed25519 signatures + blockchain anchoring
Supply chain (patterns)PatternRegistry.loadFromUrl with live update capability
Partial streaming coverage2048-char rolling window catches cross-chunk secrets
Credential rotation pre-vaultrotationHook — rotate before the secret is stored
Observability data leakOutcome-only audit schema — never the secret