Skip to content
PromptGlassDocs

Policy reference

This page documents every key in a PromptGlass Policy resource. Policies are applied with promptglass policy apply -f <file> or edited in the console. Unknown keys are rejected at apply time.

Top level

Section titled “Top level”
apiVersion: promptglass.ai/v1
kind: Policy
metadata:
  name: string            # unique, kebab-case
  description: string
  labels: { key: value }
spec:
  priority: 100           # lower runs first when several policies apply to a user
  appliesTo:              # optional; default = everyone
    groups: [string]      # IdP group names or IDs
    departments: [string]
    excludeGroups: [string]
  defaults:
    storage: full | redacted | metadata-only | hashed
    on_error: fail_open | fail_closed
    retention: 90d
  rules: [Rule]
  exceptions: [Exception]

Rule

Section titled “Rule”
- name: string                       # required, unique within the policy
  mode: enforce | shadow             # default enforce
  match: Match                       # omitted = match everything
  action: allow | log | coach | redact | block | route | alert
  message: string                    # shown for coach/block; supports {{app}}, {{data_classes}}, {{approved_app}}
  offer: { app: string, url: string }   # suggested alternative (coach/block)
  allowOverride: true                # coach only: user may continue with a reason
  redact:
    style: token | mask | remove     # [PERSON_1] | ****** | (deleted)
    restoreInResponse: true
  route:
    target: string                   # gateway target name
    keepModelFamily: true
  alert:
    channels: [siem, slack:#ai-risk, teams:AI Governance, servicenow, webhook:name]
    severity: low | medium | high | critical
    throttle: 1h                     # per user+rule
  storage: full | redacted | metadata-only | hashed   # overrides defaults
  retention: 30d
  on_error: fail_open | fail_closed

Match

Section titled “Match”

All conditions in a match must be true (AND). List values inside a condition are OR’d. Wildcards (*) are supported in names.

match:
  sensor: [gateway, extension, agent, connector]
  app: [chatgpt-personal, claude-*, "category:ai-coding"]
  appRisk: { min: high }                       # from the AI app catalog
  sanctioned: false                            # app not on your approved list
  model: [gpt-4o, "claude-*"]
  direction: prompt | response | upload | paste
  data_class: [pii.*, phi.*]
  min_confidence: low | medium | high
  min_count: 1
  topic: [legal-contract-review, hr-performance]
  file:
    types: [xlsx, csv, pdf]
    min_size: 1MB
    label: [Confidential, "Highly Confidential"]
  users: [alice@example.com]
  groups: [Finance]
  departments: [Legal]
  devices: { managed: false }
  network: { cidrs: [10.0.0.0/8], outside: false }
  time: { days: [Mon-Fri], hours: "08:00-18:00", tz: "America/Toronto" }
  budget: { scope: team, exceeded: true }
  tokens: { min: 50000 }                       # very large prompts

Exception

Section titled “Exception”

Exceptions are time-boxed waivers. They require an approver and are audited.

exceptions:
  - rule: block-source-code-public-ai
    groups: [OSS-Maintainers]
    reason: "Upstream open-source contributions"
    approvedBy: ciso@example.com
    expires: 2027-03-31

Storage modes

Section titled “Storage modes”
Mode Prompt/response stored Detections stored Searchable
full Yes, encrypted Yes Full text, permissioned
redacted Yes, with sensitive spans replaced Yes Redacted text
metadata-only No Yes (class, counts) Metadata only
hashed SHA-256 of normalized text only Yes Duplicate detection only

Validation

Section titled “Validation”
Terminal window
promptglass policy lint -f policies/          # schema + unreachable-rule warnings
promptglass policy test -f policies/ --text "…" --user bob@example.com --app chatgpt-personal
promptglass policy simulate -f policies/ --since 14d --report out.html

lint warns about rules that can never match because an earlier terminal rule always wins, and about route targets that don’t exist.