Threats and controls
Scroll horizontally to see the full comparison.
| Threat | Review or control | Gate |
|---|---|---|
| Untrusted instructions | Review instruction contents and declared behavior | Scan |
| Undeclared access | Review tool use and sandbox observations | Scan |
| Incomplete bundle analysis | Review the full package and escalate gaps | Scan |
| Incorrect output | Functional and adversarial tests for declared failure modes | Test |
| Regression | Compare behavior with the previous approved version | Test |
| Dependency drift | Re-test when a model, tool, or test case changes | Continuous |
| Changes after approval | Bind evidence and signature to a content hash | Sign |
| Version drift | Install an exact immutable release | Publish |
| Unreviewed community releases | Allow only workspace-approved trust tiers | Install |
| Misuse during execution | Apply runtime permissions and decision boundaries | Runtime |
Trust tiers and evidence
Scroll horizontally to see the full comparison.
| Control | Official | Verified Partner | Community |
|---|---|---|---|
| Publisher signature | Platform-signed | Attested partner key | Publisher signature |
| Workspace eligibility | Regulated default | Workspace opt-in | Off by default |
| Review path | Architecture review and full evaluator suite | Partner review and evaluator suite | Catalog review and scanning |
| Deprecation | 12-month notice | Standard notice | Publisher-managed |
Operating the registry
Workspace privacy
Discuss private distribution and access requirements before publishing business-specific skills.
Deployment requirements
Review hosted, customer-cloud, and on-premises requirements with the team.
Audit evidence
Use release records to connect scans, evaluations, signatures, and installation history.
Report a skill
Open the relevant listing and use its reporting controls to raise a concern.
Browse the catalog
Governance across your agent estate
Review wider agent risk and compliance controls with AgentGRC.