Skip to main content
OISP Policies define rules for how the sensor handles AI API traffic. Policies enable you to:
  • Allow: Permit requests (default behavior)
  • Block: Prevent requests from reaching AI providers
  • Redact: Remove sensitive data before forwarding
  • Alert: Allow but generate notifications for review

​
Quick Start

Create a policy file at ~/.config/oisp-sensor/policies.yaml: 
version: "1.0"
metadata:
  name: "My AI Policy"

settings:
  default_action: allow

policies:
  - id: "block-unapproved-providers"
    name: "Block Unapproved AI Providers"
    enabled: true
    priority: 100
    match:
      provider:
        not_in: [openai, anthropic, google]
    action: block
    reason: "Provider not on approved list"
Enable policies in your config: 
# ~/.config/oisp-sensor/config.toml
[policy]
enabled = true
file = "~/.config/oisp-sensor/policies.yaml"
hot_reload = true

​
Policy Structure

​
Basic Format

version: "1.0"
metadata:
  name: "Policy Name"
  description: "Description"

settings:
  default_action: allow    # allow | block
  fail_mode: open          # open | closed

policies:
  - id: "unique-id"
    name: "Human-readable name"
    enabled: true
    priority: 100          # Higher = evaluated first
    match:
      # Conditions
    action: block          # allow | block | redact | alert
    reason: "Why blocked"

​
Supported Formats

FormatUse CaseExample
YAMLCloud sync, complex policiespolicies.yaml
TOMLLocal config, developer usepolicies.toml
JSONAPI responses, programmaticpolicies.json

​
Match Conditions

All conditions in a policy are AND-ed. For OR logic, use multiple policies.

​
Provider

match:
  provider:
    equals: "openai"
    in: [openai, anthropic, google]
    not_in: [deepseek, baidu]
    matches: "openai*"    # Wildcard

​
Model

match:
  model:
    equals: "gpt-4o"
    family: "gpt-4"       # Matches gpt-4, gpt-4o, gpt-4-turbo
    in: [gpt-4o, claude-3-opus]
    not_in: [gpt-4-32k]   # Block expensive models

​
Process

match:
  process:
    name:
      equals: "cursor"
      in: [cursor, code, claude]
      matches: "*python*"
    path:
      starts_with: "/Applications/"
      matches: "*/node_modules/*"

​
Actor (User)

match:
  actor:
    user:
      equals: "john"
      in: [john, jane]
    uid:
      greater_than: 1000    # Non-system users

​
Content (for Redaction)

match:
  content:
    contains_pattern:
      - email
      - credit_card
      - ssn
      - api_key
    contains_keywords:
      - password
      - secret

​
Time-Based

match:
  time:
    hour:
      between: [9, 17]      # Business hours
    day:
      in: [monday, tuesday, wednesday, thursday, friday]
    timezone: "America/New_York"

​
Cost/Usage

match:
  cost:
    request_cost:
      greater_than: 1.00    # USD
    user_daily_total:
      greater_than: 50.00

​
Actions

​
Allow

Permit the request (default). 
action: allow

​
Block

Prevent the request from reaching the AI provider. 
action: block
reason: "Human-readable reason"

​
Redact

Modify the request to remove sensitive data. 
action: redact
redaction:
  targets: [request_body]
  patterns:
    - type: email
      replacement: "[EMAIL REDACTED]"
    - type: api_key
      replacement: "[API KEY REDACTED]"
    - type: regex
      pattern: "\\b\\d{3}-\\d{2}-\\d{4}\\b"
      replacement: "[SSN REDACTED]"
  mode: mask    # mask | hash | remove

​
Alert

Allow but generate an alert. 
action: alert
alert:
  severity: high    # low | medium | high | critical
  channels: [webhook]
  message: "Unusual AI usage detected"

​
Examples

​
Block Chinese AI Providers

policies:
  - id: "block-chinese-providers"
    name: "Block Chinese AI Providers"
    priority: 100
    match:
      provider:
        in: [deepseek, baidu, qwen, zhipu]
    action: block
    reason: "Chinese AI providers not approved"

​
Approved Apps Only

settings:
  default_action: block    # Deny by default

policies:
  - id: "allow-approved-apps"
    priority: 100
    match:
      process:
        name:
          in: [cursor, claude, copilot]
    action: allow

​
Redact PII in External Requests

policies:
  - id: "redact-pii"
    priority: 100
    match:
      provider:
        not_in: [ollama]    # External only
      content:
        contains_pattern: [email, phone, ssn]
    action: redact
    redaction:
      targets: [request_body]
      patterns:
        - type: email
          replacement: "[EMAIL]"
        - type: phone
          replacement: "[PHONE]"
        - type: ssn
          replacement: "[SSN]"

​
Cost Controls

policies:
  # Block expensive models for non-leads
  - id: "restrict-gpt4"
    priority: 100
    match:
      model:
        in: [gpt-4, claude-3-opus]
    action: block
    reason: "Premium models require approval"

  # Daily spend limit
  - id: "daily-limit"
    priority: 90
    match:
      cost:
        user_daily_total:
          greater_than: 50.00
    action: block
    reason: "Daily AI spend limit exceeded ($50)"

​
Business Hours Only

policies:
  - id: "business-hours"
    priority: 100
    match:
      time:
        hour:
          not_between: [9, 17]
    action: block
    reason: "AI access restricted to business hours"

​
Evaluation Logic

​
Order

  1. Policies sorted by priority (highest first)
  2. First matching policy wins
  3. If no match, default_action applies

​
Exception Handling

Use except to carve out exceptions: 
- id: "block-external"
  match:
    provider:
      not_in: [ollama]
  except:
    actor:
      user:
        in: [admin, researcher]
  action: block

​
Fail Mode

settings:
  fail_mode: open     # Allow if policy engine fails
  # fail_mode: closed # Block if policy engine fails

​
Built-in Patterns

PatternDescriptionExample
emailEmail addresses[email protected]
phonePhone numbers+1-555-123-4567
ssnUS Social Security123-45-6789
credit_cardCredit cards4111-1111-1111-1111
api_keyAPI keyssk-xxx, key-xxx

​
Hot Reload

Policies reload automatically when the file changes: 
[policy]
hot_reload = true
Or send SIGHUP to reload manually: 
kill -HUP $(pgrep oisp-sensor)

​
Oximy Cloud Sync

With Oximy Cloud, policies sync from the cloud: 
[mode]
connection = "oximy"

[oximy]
api_key = "oxm_live_xxx"
policy_sync_interval = 60    # seconds
Cloud policies override local policies when connected.