Skip to main content
A detector is a named, reusable detection: one capability from the detector catalog configured for your needs. It is the unit you compose into collector chains — the same detector (say “Block jailbreaks (strict)”) can be attached to many collectors. You create and edit detectors in the console (Detectors). A detector is defined by:
PropertyValuesMeaning
Typeone of the catalogThe detection capability it configures (e.g. prompt_guard).
Modeobserve · block · redactWhat a detection does.
Protocolall · llm · mcp · a2aWhich traffic type it applies to.
Directioninput · outputWhether it inspects prompts or responses.
Settingstype-specificThe detection’s configuration (thresholds, entities, lists…).

​
Mode — what a detection does

The mode is the detector’s enforcement intent. It changes how a detection shapes the guard response, not whether the detector runs.
ModeConsole labelReports findingis_flaggedRewrites payload
observeObserve✅——
blockEnforce✅✅—
redactRedact✅—✅
redact is only available for mutable detectors (today: data_loss_prevention). A typical rollout starts a detector in Observe to measure signal and false positives, then promotes it to Enforce (or Redact for DLP) once you trust it.

​
Direction — input vs output

  • Input detectors inspect what goes into the model/agent (the prompt, an uploaded document, requested tools).
  • Output detectors inspect what comes back (the completion, the tool calls the model emitted).
A detector only runs on requests whose direction matches. Many detectors support both sides; some are input-only or output-only (see each one’s supported sides in the catalog).

​
Protocol — LLM, MCP, A2A

protocol scopes a detector to a traffic type. all matches everything; llm, mcp, and a2a match that protocol only. The agent/MCP detectors are mcp-scoped, for example.

​
Settings

Each catalog type exposes a settings form (thresholds, entity lists, allow/deny lists, provider credentials, …). The console renders and validates this form for you. See the detector catalog for each type’s settings. Creating a detector doesn’t run it — attach it to a collector on its Policies tab to put it in the chain.