Security model
Trust boundaries
| Boundary | Trust posture |
|---|---|
| Agent | Untrusted. Presents a credential; never authorises itself. |
| Gateway, coordinator, identity registry, policy engine | Trusted. Decisions are made here. |
| Target system | Not observed by AWARE today. No execution verification. |
Mechanisms
| Mechanism | Algorithm |
|---|---|
| Credential storage | PBKDF2, per-credential salt, server-side pepper |
| Credential comparison | Constant-time |
| Credential generation | 32 bytes CSPRNG |
| Session token | JWT, HS256 |
| Audit chain | SHA-256 over canonical serialisation plus previous hash |
Threats AWARE addresses
- An agent calling a tool it has no policy for — denied by default.
- A revoked agent continuing to act — its credential stops verifying.
- Silent failure inside the decision path — failure outcomes are written to the chain.
- Evidence that cannot be independently checked — the verifier is decoupled from the writer.
Threats AWARE does not address
- A compromised host running the agent.
- A leaked bearer credential, before revocation is triggered.
- An action that was authorised but carried out wrongly — no execution verification.
- A target system that ignores the decision it was given.