Skip to main content

Built-in Policies

LangGuard includes 10 pre-configured policies covering security, compliance, cost, and performance governance.

Security Policies​

PII Data Detection​

Severity: Critical

Detects personally identifiable information in AI agent outputs.

What it detects:

  • Email addresses
  • Phone numbers
  • Social Security numbers
  • Credit card numbers
  • IP addresses
  • Physical addresses

Configuration:

# Patterns checked
patterns := [
    `[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}`,  # Email
    `\b\d{3}[-.]?\d{3}[-.]?\d{4}\b`,                     # Phone
    `\b\d{3}[-]?\d{2}[-]?\d{4}\b`,                       # SSN
    `\b\d{4}[-\s]?\d{4}[-\s]?\d{4}[-\s]?\d{4}\b`        # Credit card
]

When to use: Always enabled for production environments handling user data.

SQL Injection Prevention​

Severity: High

Blocks dangerous SQL statements that could modify or destroy data.

What it detects:

  • DROP statements
  • ALTER statements
  • TRUNCATE statements
  • DELETE without WHERE
  • Schema modifications

Example violation:

DROP TABLE users;  -- Blocked
ALTER TABLE accounts ADD column;  -- Blocked

When to use: Enable when agents can generate or execute SQL queries.

Prompt Injection Detection​

Severity: Critical

Detects attempts to manipulate AI agents through malicious prompts.

What it detects:

  • "Ignore previous instructions"
  • "Disregard all rules"
  • "Pretend you are..."
  • "Jailbreak" attempts
  • System prompt extraction

Patterns:

injection_patterns := [
    `(?i)ignore (all )?(previous|prior|above)`,
    `(?i)disregard (all )?(rules|instructions)`,
    `(?i)pretend (you are|to be)`,
    `(?i)forget (everything|all)`,
    `(?i)what (is|are) your (system|initial) (prompt|instructions)`
]

When to use: Always enabled for user-facing AI applications.

Compliance Policies​

Sensitive Data Access​

Severity: High

Monitors access to sensitive data categories.

What it detects:

  • Access to PII columns
  • Medical records queries
  • Financial data access
  • Authentication data exposure

Configuration:

sensitive_patterns:
  - "*password*"
  - "*ssn*"
  - "*social_security*"
  - "*credit_card*"
  - "*medical*"
  - "*health*"
  - "*salary*"
  - "*bank_account*"

When to use: Required for HIPAA, PCI-DSS, GDPR compliance.

Data Retention Compliance​

Severity: High

Ensures data handling aligns with retention policies.

What it checks:

  • Data accessed beyond retention period
  • Requests for deleted data
  • Archive compliance

When to use: Required for regulatory compliance (GDPR right to erasure, etc.).

Model Version Control​

Severity: Medium

Ensures only approved model versions are used.

What it checks:

  • Model name against allowlist
  • Model version validation
  • Deprecated model detection

Configuration:

approved_models:
  - "gpt-4"
  - "gpt-4-turbo"
  - "claude-3-opus"
  - "claude-3-sonnet"

deprecated_models:
  - "gpt-3.5-turbo-0301"
  - "text-davinci-003"

When to use: When you need to control which AI models are used.

Hallucination Risk Assessment​

Severity: Medium

Flags responses with low confidence or potential hallucinations.

What it checks:

  • Confidence scores below threshold
  • Claims without citations
  • Factual inconsistencies
  • "I don't know" overrides

Thresholds:

min_confidence: 0.7
require_citations: true
max_claims_without_source: 2

When to use: For applications where accuracy is critical (medical, legal, financial).

Cost Policies​

Token Usage Limits​

Severity: Medium

Enforces token consumption limits to control costs.

What it checks:

  • Input tokens per request
  • Output tokens per request
  • Total tokens per request
  • Daily/monthly limits (coming soon)

Default limits:

max_input_tokens: 4000
max_output_tokens: 2000
max_total_tokens: 6000

When to use: Always enabled to prevent runaway costs.

Output Length Limits​

Severity: Medium

Limits response length to control costs and ensure conciseness.

What it checks:

  • Character count
  • Word count
  • Token count

Default limits:

max_output_characters: 10000
max_output_words: 2000
max_output_tokens: 2000

When to use: For applications with response size requirements.

Performance Policies​

Rate Limiting​

Severity: Medium

Enforces request rate limits per agent or user.

What it checks:

  • Requests per minute
  • Requests per hour
  • Concurrent requests

Default limits:

max_requests_per_minute: 60
max_requests_per_hour: 1000
max_concurrent: 10

When to use: To prevent abuse and ensure fair resource allocation.

Enabling Policies​

Via UI​

  1. Navigate to Policies
  2. Find the policy
  3. Toggle Enabled

Via API​

POST /api/policies/{id}/toggle
Content-Type: application/json

{
  "enabled": true
}

Customizing Built-in Policies​

You can customize built-in policies by editing their configuration:

  1. Click the policy to open details
  2. Click Edit
  3. Modify thresholds or patterns
  4. Click Save
note

Customizations are applied immediately to new traces. Existing violations are not re-evaluated.

Policy Evaluation Order​

When multiple policies could apply to a trace:

  1. All enabled policies are evaluated
  2. Each policy runs independently
  3. All violations are recorded (no short-circuit)
  4. Violations are sorted by severity for display

Next Steps​