Dashboard
Content policies
Add your own phrases, values and documents that AgentID should allow, alert on, mask or block.
Built-in detectors find personal data and secrets. Content policies cover what only you know is sensitive: project code names, client names, internal identifiers or confidential documents.
Where to manage them
| Scope | Location | Applies to |
|---|---|---|
| Organization | Content Policies in the main navigation, Shared organization rules | Every AI system in the organization |
| AI system | System → Content Policies (admins) | That system only |
Policy resolution order: AgentID uses the custom policies attached directly to an AI system when they exist; otherwise it uses the shared organization policy set. Keep emergency overrides system-specific so one pilot cannot affect all systems.
Enforcement modes
| Mode | Effect on a matching prompt |
|---|---|
| Allow | Treats the phrase as safe, so built-in detectors do not mask or block it |
| Alert only | Lets the prompt through and records a security alert |
| Mask | Replaces the phrase before the prompt reaches the model |
| Block | Stops the prompt with the message Policy violation: Term blocked by admin. |
Policy values are stored protected; the dashboard shows only a preview of each phrase.
Adding policies
- Add Content Policy: enter an exact phrase or sensitive value (for example
Project Apollo) and choose the mode. - Import Content Policies: upload phrases in bulk from a
.csvfile. - Pattern Catalog Bulk Import: upload a large pattern catalog from
.csv,.jsonor.ndjson; the import streams and shows its progress. - Test Policy Pack: paste a test prompt, choose Enforce or Shadow, and click Run Test to see the decision and which signals matched.
Bulk actions let you deactivate all active policies, restore inactive ones, or permanently purge all policies (type PURGE to confirm). Import and write actions depend on your role.
Document policies
Under Document Policies, Upload Policy Document adds a confidential document and a short description of what it covers. AgentID extracts its text and checks prompts against it; a prompt that reproduces a blocked document is stopped with Policy violation: Document blocked by admin. Source files are stored privately; text extraction runs as a background job after upload.