Policy-action catalog

When a request is classified, ThreatLens looks up the policy matrix cell for that data class and resolves an action. This page is the authoritative list of what each action does.

Actions

Listed least to most restrictive.

Action	Label	What happens
allow	Allow (send as-is)	The request goes to the model unchanged.
redact	Redact sensitive values, then send	Sensitive spans are masked, then the redacted request is sent. The model never sees the raw values.
route	Route to an approved destination (else block)	The request is sent only to a destination that meets the required trust tier. If no such destination exists, it is blocked.
warn	Warn the user and send	The user sees a governance warning, then the request proceeds.
approval	Require admin approval	The request is held until an approver allows or denies it (exceptions).
block	Block	The request is stopped. It never reaches any model.

Route is the workhorse for confidential business data

For most High/Critical classes, the default is route — the content is allowed, but only to a destination you trust (your BYOK enterprise endpoint), never to a public frontier model.

Internet access modes

Independently of the action, each class also carries an internet policy that governs whether the request may use live web/grounding:

Mode	Label	Meaning
allow	Allow web (as-is)	The request may use web search/grounding normally.
safe_only	Web via safe query only	A scrubbed, sensitive-data-free query may go to the web — the raw content does not.
block	No web access	The request may not touch the internet.

Always-on protections

Two protections are not matrix-tunable — they apply on every tier, regardless of the cell you set:

• Raw secrets and credentials (the Source Code / Secrets class) are always blocked.
• Prompt-injection patterns are always blocked before the request reaches a model.

How the action is chosen

1. Content is classified into a data class.
2. The matrix cell for that class yields a minimum trust tier, an action, and an internet mode.
3. The chosen destination is checked against the minimum trust tier. If it qualifies, the action runs as configured; if it does not, route falls back to block.
4. The outcome — including the action, the destination actually used, and any redactions — is written to the audit record.

In monitor mode, the action is computed and recorded as the intended decision but not enforced; in enforce mode it is applied. See monitor vs enforce.