You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Cryptographic enforcement mitigations for AAI001-AAI008 via Agent Passport System #13
The Agentic AI Top 10 maps well to the Agent Passport System's enforcement model. Here's how APS addresses each relevant risk:
Risk
APS Mitigation
AAI001: Agent Hijacking
Delegation chain check runs in enforcement layer, not model layer. Hijacked agent's intent is signed (forensic evidence) then denied if out-of-scope
AAI002: Tool Misuse
ProxyGateway binds exact tool + parameters + target at approval time. Tool calls outside approved scope are rejected before execution
AAI003: Privilege Escalation
Monotonic narrowing invariant — each delegation can only narrow authority, never widen. Mathematically impossible for sub-agent to exceed parent's scope
AAI005: Agent Impersonation
Ed25519 cryptographic identity per agent. Messages must be signed. Unsigned or forged messages fail signature verification
AAI006: Memory/Context Poisoning
3-signature receipt chain provides tamper-proof audit trail. Poisoned context can cause agent to request unauthorized action, but policy engine denies it
AAI008: Agent Orchestration Attacks
Cascade revocation — revoking one delegation invalidates all downstream delegations instantly, severing any compromised orchestration chain
Key design principle: APS treats these as authorization enforcement problems, not model behavior problems. The enforcement layer operates independently of the LLM, so compromised model reasoning cannot bypass delegation checks.
The protocol is open source (Apache 2.0), ships as both TypeScript and Python SDKs, and has an MCP server (120 tools) for interoperability with any agent framework.
Activity