Threat Signals & Approvals
Turn block telemetry and coverage gaps into guard proposals. Close the loop from live blocks back to published guards. Human-reviewed publish updates compileHash across every surface.
This page is the improve chapter in Operate. After you prove enforcement in Audit & metrics, use Threat signals and Rule approvals to close the loop from live blocks back to published guards. Nothing applies silently: proposals stage into the same runtime policy bundle you author on Runtime policy.
The improve loop
Recognition and improvement share one spine. Typed patches from Discover hand into execution policy compile via ThreatSignalPatch (capability ids, optional MCP tool deltas, suggestion id).
- Telemetry. Blocks and metadata from enforce, MCP, SDK, and cells.
- Threat signals. Queue ranked patterns with severity and suggested capabilities.
- Rule approvals. Human replay on recent samples before any staging.
- Runtime policy. Publish produces a new compileHash for all surfaces.
- Verify. Runs and audit log close the loop.
Protect rail map
Recognize patterns, approve changes, then publish.
Recognize
- Threat signalsUncovered blocks, catalog gaps, and fleet quorum items.
- Audit logBlock and mask metadata that feeds recognition.
Approve
- Rule approvalsReplay, diff, approve, stage, and dismiss false positives.
Publish
- Runtime policyPublish one compiled bundle. Every surface inherits compileHash.
- RunsVerify new act rows after publish.
Threat signals
Open Protect → Threat signals for the ranked queue. Filter by Live blocks, Threat (catalog gap), or Fleet quorum. KPIs summarize uncovered patterns, seven-day blocks, catalog threats, and pending fleet contributions.
| Type | Source | Typical next step |
|---|---|---|
Live blocklive_block | Block outcomes from governed paths with no matching deny in workspace guards. | Draft in Runtime policy or open Rule approvals. |
Threatcatalog_gap | High or critical catalog incidents whose attack surfaces are not covered by guards. | Review proposal in Rule approvals. |
Fleet quorumfederated_pending | Pending anonymized federated contributions before fleet pack promotion. | Rule approvals and fleet pack review when signing is enabled. |
Rows include severity, attack surfaces, and suggested capabilities. Use Draft in Runtime policy for live blocks, Review proposal for catalog gaps, and Rule approvals for fleet quorum items.
Rule approvals
Open Protect → Rule approvals. Use Synthesize when new blocks appear and you want a fresh proposal from telemetry. The queue shows pending items with replay metadata before you approve.
| Action | Effect |
|---|---|
| Synthesize | Generate a guard proposal from current telemetry. May return no new rules. |
| Approve and stage | Record the decision and stage changes for Runtime policy publish. |
| Stage for publish | Move the diff into the publish flow without skipping human review. |
| Dismiss as false positive | Reject the proposal so repeated noise can be suppressed. |
| Open in Runtime policy | Deep link with suggestion id and capability configure context. |
Each proposal includes a JSON diff (typically MCP tool policy deltas), simulation outcome, and a deep link to Runtime policy with suggestionId. Integrators can preview typed patches via the workspace rule-approvals apply API before publish.
When Fleet pack signing is enabled, approved rules can ship in a versioned signed pack. Otherwise the console shows an honest empty state until federated signing is configured.
Publish and verify
After approve and stage, open Runtime policy to review the change set and publish. Follow the verify steps from Runtime policy (API, SDK, or policy stream).
- New runs show the updated compileHash on the run row.
- Audit log filtered by that hash shows the expected verdict mix.
- Live block signals quiet or move to dismissed after the guard matches telemetry.
- Use simulatePolicy() before widening enforce on high-risk tools.
How recognition works
Live recognition groups recent block outcomes where workspace guards do not explain the deny. Catalog recognition compares high-severity incident surfaces against your published MCP tool policy. Fleet recognition (when enabled) surfaces quorum work for anonymized federated contributions before pack promotion.
This is not autonomous policy rewrite: recognition proposes typed changes, humans approve, and publish updates compileHash for every connected surface. Editorial catalogs alone do not close your loop; your own blocks and audit metadata do.
Roles and access
| Surface | Requirement |
|---|---|
| Threat signals | Platform Team workspace (hosted act and enforce traffic). |
| Rule approvals | Platform Team workspace. Approve and dismiss need admin or owner. |
| Fleet pack | Optional signed pack when federated signing is enabled for the workspace. |