SafeCadencePRE-AUTHORIZED
DO
Analyze approved data. Correlate findings. Prepare a recommendation within the mission.
AI agent security & control / Preview
AgentGuard
Know what agents do. Control what they may do.
Keep business agents within their authorized mission. AgentGuard is the policy, approval and evidence layer for execution through supported runtimes and tool gateways.
AI at Scale. Humans in Control.
The authority model
DO. NOD. NEVER.
Selected guarded paths are tested locally. Enterprise deployment needs separate qualification.
Coverage depends on actions passing through the guarded runtime or gateway. Recording an action is not the same as enforcing permission.
Covered execution boundary
Permission before action. Evidence after.
Locally tested admissions and fictional callbacks, not a customer production demonstration.
- Propose
Bind an action to the agent, mission, tool and target.
- Decide
Evaluate scope and policy before tool execution.
- Approve or block
Hold consequential work for authorized approval. Deny work outside scope.
- Verify & record
Capture the result, evidence and failures without private chain-of-thought.
A history people can review
Supported records link the proposed operation, principal, target digest, policy decision, scoped grant, result and original evidence chain. Actual record coverage varies by path. Sensitive raw results are not a substitute for reviewable evidence.
A returned callback is not independent outcome verification. Exclusive effect holds are not automatically released by callback completion; unsupported reconciliation remains held.
What does stopping an agent mean?
Stop and revocation controls must prevent future actions through the guarded path. They do not undo completed actions or control tools running outside that path. Enterprise stop behavior remains subject to validation.
Capability transparency
What is ready today?
Reviewed October 14, 2026. Local tests and fixtures are not real-environment certification.
| Capability | Status | What this means today |
|---|---|---|
| Policy and scoped admissions | Tested locally | Supported local paths test allow/ask/deny, principal and target scope, expiry and consumption limits before callback execution. This is not an OS sandbox. |
| Approvals, stops and revocation | Tested locally | Local approval holds and future-action stop/revocation paths are tested. They do not undo completed effects or control unguarded tools. |
| Original evidence integrity | Tested locally | Original SDK journals are read back and checked in local acceptance. Integrity checks detect covered changes, not every privileged deletion or external outcome. |
| Independent outcome verification | Bounded preview | Verification is profile-specific. A callback returning or a valid journal does not prove business success, exactly-once delivery or rollback. |
Start with a bounded business pilot.
Agree on the mission, data access, owner and acceptance tests before connecting systems.