Policies Overview

LangGuard's policy engine enables automated governance and compliance for AI operations. Define rules, detect violations, and ensure your AI systems meet organizational standards.

What Are Policies?

Policies are rules that evaluate AI agent traces to detect:

• Security risks - PII exposure, credential leaks, unapproved tools, identity violations
• Compliance violations - Model approval, cross-boundary access, audit requirements
• Cost overruns - Budget thresholds, per-trace cost limits, token limits
• Operational health - Latency thresholds, error rates, metadata completeness

How It Works

LangGuard uses Open Policy Agent (OPA) with policies written in Rego:

Trace Ingested → Policy Evaluation → Violations Logged → Webhooks Triggered

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│   New Trace     │────▶│   OPA Server    │────▶│   Violations    │
│   Arrives       │     │   (Rego Eval)   │     │   Recorded      │
└─────────────────┘     └─────────────────┘     └─────────────────┘

OPA Engine Health

The Policies page displays an OPA engine health status indicator showing whether the policy evaluation engine is running correctly. A healthy status means policies are being evaluated in real time as traces arrive. If the engine is unhealthy, policy evaluations are queued and processed once the engine recovers.

Policy Evaluation Pipeline

When a new trace is ingested:

1. Trace received — LangGuard receives the trace from a connected integration
2. Policy selection — All enabled policies are identified
3. OPA evaluation — Each enabled policy's Rego code is evaluated against the trace data
4. Violation recording — Any violations are recorded with severity, evidence, and the originating policy
5. Notifications — Configured webhooks are triggered for violations

Policy Components

Policy Definition

Each policy includes:

Component	Description
Name	Unique identifier
Description	What the policy detects
Category	Security & Access, Models & Compliance, Budget & Operations, Audit
Severity	Critical, High, Medium, Low
Rego Code	The evaluation logic
Mode	Enforce, Permissive, or Disabled
Enabled	Active or disabled

Example Policy

package langguard.pii_detection

import rego.v1

# Emit a violation for each PII match when the detect_pii tag is present
violation contains result if {
    has_pii_tag
    pii := object.get(input, "pii_detection", {})
    pii.found == true
    some match in pii.matches

    result := {
        "type": "pii_detected",
        "pii_type": match.type,
        "location": match.location,
        "message": sprintf("PII detected (%s) in %s", [match.type, match.location]),
    }
}

Built-in Policies

LangGuard includes pre-configured policies across all categories:

Policy	Category	Severity	Default
Credential Surface Discovery	Security & Access	Critical	Enabled
PII Detection	Security & Access	Critical	Enabled
Non-Human Identity Violation	Security & Access	High	Disabled
Unapproved Tool Use	Security & Access	High	Enabled
Approved Model Version	Models & Compliance	High	Disabled
Latency/Health Threshold	Models & Compliance	Medium	Disabled
Cross-Boundary Access	Models & Compliance	Critical	Enabled
Budget Overrun Detection	Budget & Operations	High	Disabled
Mandatory Trace Logging	Audit	Low	Enabled
Metadata Tagging Requirements	Audit	Low	Enabled

Documentation

• Built-in Policies - Details on all pre-configured policies
• Creating Policies - Write custom policies with Rego
• Policy Violations - Understanding and managing violations

Policy Modes

Each policy can operate in one of three modes:

Mode	Behavior
Enforce	Violations are logged and the action is blocked. Use for critical policies that must prevent harmful behavior.
Permissive	Violations are logged but the action is allowed to proceed. Use for monitoring and tuning new policies before enforcing them.
Disabled	The policy is not evaluated. No violations are recorded.

Choosing a Mode

• Start with Permissive — When deploying a new policy, use Permissive mode first to observe what it catches without impacting operations
• Promote to Enforce — Once you're confident the policy is correctly calibrated, switch to Enforce mode
• Use Disabled sparingly — Only disable policies you no longer need; consider Permissive mode instead for policies you want to pause temporarily

Best Practices

1. Start with Built-in Policies

Enable built-in policies first to establish baseline governance:

• Enable "Credential Surface Discovery" and "PII Detection" for security
• Enable "Cross-Boundary Access" to enforce environment isolation
• Enable "Budget Overrun Detection" to control costs

2. Tune Severity Levels

Adjust severity based on your risk tolerance:

• Critical: Must be addressed immediately
• High: Review same day
• Medium: Review weekly
• Low: Informational only

3. Review Violations Regularly

Establish a violation review process:

• Daily: Check critical violations
• Weekly: Review all violations
• Monthly: Analyze trends, adjust policies

4. Create Custom Policies

After understanding built-in policies, create custom ones for:

• Company-specific compliance requirements
• Application-specific guardrails
• Industry regulations

---

Next Steps

• Built-in Policies - Explore available policies
• Creating Policies - Write your own
• Troubleshooting - Get help with common issues