AGENT GUARD RUNTIME
Protect the world from your agent.
The model can reason about actions. Only the external Guardian can authorize and execute them.
-
UNTRUSTEDAgent / model processProposes actions
↓
TRUSTED ENFORCEMENTAgent Guard KernelIdentity · mission · capability · replay · risk · revocation↓
{['MCP','TOOLS','SHELL','FILES','NETWORK','SECRETS','DELEGATION','CLOUD','PAIP'].map((item) => {item})}
+
UNTRUSTEDAgent / model processProposes actions
↓
TRUSTED ENFORCEMENTAgent Guard KernelIdentity · mission · capability · replay · risk · revocation
↓
{['MCP','TOOLS','SHELL','FILES','NETWORK','SECRETS','DELEGATION','CLOUD','PAIP'].map((item) => {item})}
01Capability control
Short-lived, mission-bound, revocable authority instead of raw credentials.
02Delegation guard
Every child remains a strict subset of parent scope, time, risk, and depth.
03Blind secrets
Perform authenticated work without revealing credentials to the model.
04Replay defense
Reject reused request IDs, nonces, and stale authorization envelopes.
05Quarantine trees
Stop an agent and its known delegated descendants immediately.
06Flight recorder
Hash-chain structured actions without collecting private chain-of-thought.