Policies

The Policies page is where you define what the scoring engine looks for. It has three tabs (Policies, Host Detection and Scoring Config) with the policy list as the default view.

The policy list

A summary strip shows totals at a glance (e.g. 20 Policies, 20 Active, 0 Disabled, 12 Critical, 8 High). Each policy is a row (examples: agent_restrictive_tool_calling, sql_injection, common_online_racism_speech, exfil) showing type, status, severity, performance and last-modified. Add Policy opens the editor.

Adding or editing a policy

Field	Meaning
Policy Title	The policy's name
Description	What it is for
Keywords	Comma-separated terms, pushed to the FPGA for hardware-accelerated matching
Type	LLM Security · Code Security · Content Filter · Data Privacy
Status	Active · Testing · Disabled
Severity	None · High · Critical

Use Testing status to trial a policy. It records matches without enforcing. Then promote it to Active once you have confirmed its behaviour in the Security Lab.

Tip. Keywords are compiled to the FPGA, so large keyword sets stay fast. After adding a policy, confirm it in the Security Lab with a prompt designed to trip it.

Host Detection

Governs detection rules tied to hosts/endpoints rather than prompt content. Use it to tune which hosts are watched and how their activity is treated. (To be expanded in a future manual revision.)

Scoring Config

Holds the engine-level settings that turn individual detections into a composite score and a combined action: the thresholds and weighting that determine where SAFE becomes MONITOR and MONITOR becomes BLOCK. Changing these affects every verdict, so adjust deliberately and re-test in the Security Lab.

Access

Creating, editing, enabling, disabling and deleting policies, and changing scoring configuration, are Administrator-only. Viewers can read the list and open a policy but cannot save.

See also: Policy store file format.