Policies overview

A PromptGlass policy is an ordered list of rules. Each rule matches on who (users, groups, departments), where (AI app, model, sensor), what (data classes, topics, file types) and how much (budgets, volume). It then takes one action. Policies are YAML, version-controlled, and evaluated in microseconds at the gateway and in the browser extension.

Actions

Action	Effect	Available on
allow	Pass through and record metadata	All sensors
log	Pass through and record per the storage mode	All sensors
coach	Show the user a message; they can switch tools, edit, or continue with a reason	Extension, gateway (web)
redact	Replace sensitive spans with tokens before sending; optionally restore in the response	Extension, gateway
block	Stop the request and show a message	Extension, gateway
route	Send the request to an approved model or tenant instead	Gateway (API and supported web apps)
alert	Pass through and notify (SIEM, Slack/Teams, ticket)	All sensors, including connectors

Connectors and the endpoint agent observe after the fact, so on those sensors any rule that would coach, redact, block or route is recorded as a would-have action and can alert.

Evaluation order

Rules are evaluated top-down, and the first matching rule with a terminal action wins (allow, coach, redact, block, route). log and alert are non-terminal: they record and notify, then evaluation continues. That lets you put broad logging rules first and enforcement rules after them.

A minimal policy

apiVersion: promptglass.ai/v1
kind: Policy
metadata:
  name: baseline
spec:
  defaults:
    storage: redacted
    on_error: fail_open
  rules:
    - name: block-secrets-everywhere
      match: { data_class: [secret.*] }
      action: block
      message: "This looks like a credential. Remove it before using AI. Need help? #security"

    - name: coach-personal-chatgpt
      match: { app: [chatgpt-personal, gemini-personal] }
      action: coach
      message: "Please use ChatGPT (Acme workspace) for work. It keeps your data out of model training."
      offer: { app: chatgpt-acme, url: "https://chatgpt.com/?workspace=acme" }

    - name: log-everything
      action: log

Rollout pattern

1. Observe: start every new rule as mode: shadow. It records the action it would have taken without affecting users.
2. Coach: switch to coach and watch the override rate. A high override rate means the rule is too broad or the approved alternative is lacking.
3. Enforce: move to redact or block for the data classes where the risk justifies it.

- name: redact-customer-records
  mode: shadow          # shadow | enforce
  match: { data_class: [acme.customer_record] }
  action: redact

Managing policies

• Console: a visual editor with a live test panel.
• CLI / GitOps: promptglass policy apply -f policies/ from CI. The console shows Git-managed policies as read-only.
• Simulation: promptglass policy simulate --since 7d replays recent events against a draft policy and reports how its actions would differ.

See the policy reference and examples.