Policy Examples

This guide provides practical policy examples for common use cases. Each example includes the policy definition, explanation, and expected behavior.

PII Protection Policies

Block SSN in Prompts

Detect and block Social Security Numbers in user prompts:

policies:
  - id: block-ssn
    name: Block SSN in Input
    description: Prevent SSN from being sent to LLMs
    type: pii
    trigger: request
    pattern: '\b(\d{3})[- ]?(\d{2})[- ]?(\d{4})\b'
    action: block
    severity: critical
    message: "Social Security Numbers cannot be processed. Please remove the SSN and try again."

Matches:

123-45-6789
123 45 6789
123456789

Does not match:

12-345-6789 (wrong grouping)
1234-56-789 (wrong format)

Redact Credit Card Numbers

Automatically redact credit card numbers in responses:

policies:
  - id: redact-credit-card
    name: Redact Credit Cards
    description: Redact credit card numbers in LLM responses
    type: pii
    trigger: response
    patterns:
      - '\b4[0-9]{12}(?:[0-9]{3})?\b'          # Visa
      - '\b5[1-5][0-9]{14}\b'                   # Mastercard
      - '\b3[47][0-9]{13}\b'                    # Amex
      - '\b6(?:011|5[0-9]{2})[0-9]{12}\b'       # Discover
    action: redact
    replacement: "[CARD-REDACTED]"
    severity: critical

Email Address Handling

Log but don't block email addresses:

policies:
  - id: log-email
    name: Log Email Addresses
    description: Log when email addresses are processed
    type: pii
    trigger: both
    pattern: '\b[a-zA-Z0-9._%+\-]+@[a-zA-Z0-9.\-]+\.[a-zA-Z]{2,}\b'
    action: log
    severity: medium
    metadata:
      category: "contact_info"
      compliance: ["GDPR", "CCPA"]

Indian PII Detection

Detect PAN and Aadhaar numbers:

policies:
  # PAN Number
  - id: redact-pan
    name: Redact Indian PAN
    description: Redact Permanent Account Numbers
    type: pii
    trigger: both
    pattern: '\b[A-Z]{3}[PCHABGJLFT][A-Z][0-9]{4}[A-Z]\b'
    action: redact
    replacement: "[PAN-REDACTED]"
    severity: critical

  # Aadhaar Number
  - id: redact-aadhaar
    name: Redact Aadhaar
    description: Redact Aadhaar unique IDs
    type: pii
    trigger: both
    pattern: '\b[2-9][0-9]{3}\s?[0-9]{4}\s?[0-9]{4}\b'
    action: redact
    replacement: "[AADHAAR-REDACTED]"
    severity: critical

SQL Injection Prevention

Block SQL Injection Patterns

Prevent SQL injection attempts in prompts:

policies:
  - id: sql-injection-union
    name: Block UNION SQL Injection
    type: security
    trigger: request
    pattern: 'union\s+select'
    action: block
    severity: critical
    message: "Potential SQL injection detected. Request blocked for security."

  - id: sql-injection-comment
    name: Block SQL Comment Injection
    type: security
    trigger: request
    pattern: '--|\*\/|\/\*'
    action: block
    severity: critical

  - id: sql-injection-always-true
    name: Block Always-True Conditions
    type: security
    trigger: request
    pattern: "1\\s*=\\s*1|''\\s*=\\s*''|\"\"\\s*=\\s*\"\""
    action: block
    severity: critical

Block Dangerous SQL Commands

Prevent destructive SQL operations:

policies:
  - id: block-drop-table
    name: Block DROP TABLE
    type: security
    trigger: request
    pattern: 'drop\s+table'
    action: block
    severity: critical
    message: "DROP TABLE commands are not permitted."

  - id: block-truncate
    name: Block TRUNCATE
    type: security
    trigger: request
    pattern: 'truncate\s+table'
    action: block
    severity: critical

  - id: block-delete-all
    name: Block DELETE without WHERE
    type: security
    trigger: request
    pattern: 'delete\s+from\s+\w+\s*(?:;|$)'
    action: block
    severity: high
    message: "DELETE without WHERE clause is not permitted."

Access Control Policies

Admin-Only Resources

Restrict access to admin tables:

policies:
  - id: admin-users-table
    name: Restrict Users Table Access
    type: access_control
    trigger: request
    pattern: '\busers\b'
    action: check_permission
    required_permission: admin
    deny_message: "Access to users table requires admin privileges."
    severity: high

  - id: admin-audit-logs
    name: Restrict Audit Log Access
    type: access_control
    trigger: request
    pattern: 'audit_log'
    action: check_permission
    required_permission: audit_read
    severity: high

Role-Based Content Filtering

Filter content based on user role:

policies:
  - id: financial-data-access
    name: Financial Data Access Control
    type: access_control
    trigger: response
    conditions:
      - user_role: not_in ["finance", "executive", "admin"]
    patterns:
      - 'revenue|profit|salary|compensation'
    action: redact
    replacement: "[FINANCIAL-DATA-RESTRICTED]"
    severity: medium

Content Moderation Policies

Block Inappropriate Content

Prevent inappropriate content in responses:

policies:
  - id: block-profanity
    name: Block Profanity
    type: content
    trigger: response
    pattern_file: /etc/axonflow/profanity-list.txt
    action: redact
    replacement: "[CONTENT-FILTERED]"
    severity: medium

  - id: block-harmful-instructions
    name: Block Harmful Instructions
    type: content
    trigger: response
    patterns:
      - 'how to (make|create|build).*(bomb|weapon|explosive)'
      - 'instructions for (illegal|harmful)'
    action: block
    severity: critical
    message: "This response has been blocked due to content policy violation."

Competitor Mention Policy

Handle competitor mentions appropriately:

policies:
  - id: competitor-mentions
    name: Flag Competitor Mentions
    type: content
    trigger: response
    patterns:
      - 'competitor_a|competitor_b|competitor_c'
    action: flag
    severity: low
    metadata:
      review_required: true
      category: "competitive_intelligence"

Rate Limiting Policies

Token Limit per Request

Limit tokens per individual request:

policies:
  - id: request-token-limit
    name: Request Token Limit
    type: rate_limit
    trigger: request
    limit:
      type: tokens
      max: 4000
    action: block
    message: "Request exceeds maximum token limit of 4000."

User Rate Limiting

Limit requests per user:

policies:
  - id: user-rate-limit
    name: User Rate Limit
    type: rate_limit
    trigger: request
    limit:
      type: requests
      max: 100
      window: 1h
      key: user_id
    action: block
    message: "Rate limit exceeded. Please try again later."

Organization Token Budget

Enforce organization-wide token budgets:

policies:
  - id: org-daily-budget
    name: Organization Daily Token Budget
    type: rate_limit
    trigger: request
    limit:
      type: tokens
      max: 1000000
      window: 24h
      key: organization_id
    action: block
    severity: high
    message: "Daily token budget exceeded for your organization."

Compliance Policies

Ensure responses don't include unnecessary personal data:

policies:
  - id: gdpr-data-minimization
    name: GDPR Data Minimization
    type: compliance
    trigger: response
    conditions:
      - request_type: not_in ["kyc", "identity_verification"]
    patterns:
      - 'date of birth|DOB|birth date'
      - 'passport number|passport no'
      - 'driver.?s? license'
    action: redact
    replacement: "[DATA-MINIMIZED]"
    severity: medium
    metadata:
      regulation: "GDPR"
      article: "5(1)(c)"

Financial Disclaimer

Add disclaimer to financial advice:

policies:
  - id: financial-disclaimer
    name: Financial Advice Disclaimer
    type: compliance
    trigger: response
    conditions:
      - response_contains: ["invest", "stock", "financial advice", "buy", "sell"]
    action: append
    append_text: "\n\n---\n*Disclaimer: This is not financial advice. Please consult a licensed financial advisor before making investment decisions.*"
    severity: medium

Audit Logging Policy

Ensure all requests are logged:

policies:
  - id: audit-all-requests
    name: Audit All Requests
    type: audit
    trigger: both
    action: log
    log_fields:
      - timestamp
      - user_id
      - organization_id
      - request_type
      - model_used
      - token_count
      - response_time_ms
      - policy_violations
    retention_days: 365

Conditional Policies

Time-Based Policy

Apply different policies based on time:

policies:
  - id: after-hours-restriction
    name: After Hours Restriction
    type: conditional
    trigger: request
    conditions:
      - time_of_day: outside ["09:00", "18:00"]
      - day_of_week: in ["Monday", "Tuesday", "Wednesday", "Thursday", "Friday"]
    action: log
    severity: low
    metadata:
      alert: "After-hours request"

  - id: weekend-restriction
    name: Weekend Access Restriction
    type: conditional
    trigger: request
    conditions:
      - day_of_week: in ["Saturday", "Sunday"]
      - user_role: not_in ["admin", "on-call"]
    action: block
    message: "System access is restricted on weekends."

Model-Specific Policy

Apply policies based on model selection:

policies:
  - id: gpt4-only-for-premium
    name: GPT-4 Premium Only
    type: conditional
    trigger: request
    conditions:
      - model: starts_with "gpt-4"
      - user_tier: not_in ["premium", "enterprise"]
    action: block
    message: "GPT-4 models are only available for premium users."

Combining Policies

Policy Groups

Group related policies:

policy_groups:
  - id: pii-protection
    name: PII Protection Suite
    policies:
      - block-ssn
      - redact-credit-card
      - redact-pan
      - redact-aadhaar
      - log-email

  - id: security-hardening
    name: Security Hardening
    policies:
      - sql-injection-union
      - sql-injection-comment
      - sql-injection-always-true
      - block-drop-table
      - block-truncate

Policy Inheritance

Extend base policies:

policies:
  - id: base-pii-policy
    name: Base PII Policy
    type: pii
    action: redact
    severity: high
    abstract: true  # Can't be used directly

  - id: healthcare-pii
    name: Healthcare PII Policy
    extends: base-pii-policy
    patterns:
      - medical_record_number
      - diagnosis_code
      - prescription_id
    severity: critical  # Override

Best Practices

Policy Naming

Use clear, descriptive names:

block-ssn - Good (clear action and target)
policy1 - Bad (not descriptive)

Severity Levels

Level	Use Case
`critical`	Security vulnerabilities, highly sensitive PII
`high`	Sensitive data, access control
`medium`	Privacy concerns, compliance
`low`	Informational, audit

Testing Policies

Always test policies before production:

# Test a specific policy
axonctl policy test --policy block-ssn --input "My SSN is 123-45-6789"

# Expected output:
# Policy: block-ssn
# Match: true
# Action: block
# Matched text: "123-45-6789"

PII Protection Policies​

Block SSN in Prompts​

Redact Credit Card Numbers​

Email Address Handling​

Indian PII Detection​

SQL Injection Prevention​

Block SQL Injection Patterns​

Block Dangerous SQL Commands​

Access Control Policies​

Admin-Only Resources​

Role-Based Content Filtering​

Content Moderation Policies​

Block Inappropriate Content​

Competitor Mention Policy​

Rate Limiting Policies​

Token Limit per Request​

User Rate Limiting​

Organization Token Budget​

Compliance Policies​

GDPR Data Minimization​

Financial Disclaimer​

Audit Logging Policy​

Conditional Policies​

Time-Based Policy​

Model-Specific Policy​

Combining Policies​

Policy Groups​

Policy Inheritance​

Best Practices​

Policy Naming​

Severity Levels​

Testing Policies​

Related​