Skip to content

Agent security

Gate and audit
agents at scale.

Opaque gives platform and security teams control over AI agent actions. Define permitted operations, require human approval, and record what happened.

Try the demo Set up locally →

Open-source core · macOS and Linux · v0.6.0 and installation

Hosted demo uses fictional data. Demo guide

A gated staging release

Configure the release workflow and policy, then prepare a manifest.

Example commands · replace <task-id>
opaque task plan --manifest ./release.json
opaque task run <task-id>
opaque task show <task-id>

Approval covers one build, one destination, and one dispatch attempt. A timeout can leave the result unknown; the attempt stays consumed. Dispatch acceptance is not deployment success.

Task setup and revocation →
Gate access.
Deny by default. Limit operations by identity and target, and require human approval before sensitive execution.
Policy rules →
Bound the work.
Give tasks an exact scope and expiry. Revoke future dispatches; an approval cannot expand the reviewed task.
Task limits →
Audit actions.
Inspect decisions, approvals, and observed outcomes. Verify exported audit ranges with independently held checkpoints.
Evidence verification →

Manage policy across brokers.

Public core v0.6.0 includes signed policy distribution and SIEM export. Fleet management is separate enterprise software.

Signed policies and audit export →

Protect the broker.

Route agent actions through Opaque and run the broker under a separate OS identity. Same-user installs do not isolate custody; direct agent access remains outside the gate.

Deployment boundaries and evidence →