Risk Assessment | Agent Inspector

Overview

Agent Inspector provides multi-dimensional risk assessment that goes beyond traditional application security. Because AI agents are non-deterministic and make autonomous decisions, they require a new approach to security and risk management.

The risk assessment capability automatically analyzes your agent across five critical dimensions, identifying security hazards, compliance gaps, and operational risks before they reach production.

Security and Behavioral Assessment

Agent Inspector performs comprehensive security analysis that identifies both traditional security vulnerabilities and behavioral risks unique to AI agents. This assessment includes checks across resource management, behavioral stability, environment configuration, and privacy compliance.

Each finding is categorized by severity with specific remediation guidance:

Critical (shown as FAIL): Security hazards that must be addressed before production
Warning (shown as WARN): Issues that may cause problems but don't block deployment
Info: Recommendations for optimization and best practices

Comprehensive security assessment showing hazards by severity and behavioral metrics

Environment & Supply Chain

Track which LLM models your agent uses, whether they're pinned to specific versions, and the security posture of your tool dependencies. Supply chain visibility is critical for production readiness and compliance.

Model version tracking, tool adoption rates, and supply chain consistency

Key Insights

Model Pinning: Identifies unpinned model versions that could introduce unexpected behavior changes
Tool Adoption: Shows which tools are actively used vs. configured but unused
Version Consistency: Detects version drift across different agent sessions
Dependency Health: Tracks third-party integrations and their usage patterns

Why Model Pinning Matters

Unpinned models (e.g., "claude-3-5-sonnet-latest") can change behavior without warning when providers update their models. In production, this creates unpredictable behavioral drift. Always pin to specific versions (e.g., "claude-3-5-sonnet-20241022") for stability.

Privacy & PII Detection

Agent Inspector automatically scans all prompts, tool calls, and responses for sensitive information using Microsoft Presidio. This helps you identify privacy risks before they become compliance violations.

Privacy & PII Compliance report showing detection findings and exposure rates

Detected Entity Types

Personal Identifiers

Names, email addresses, phone numbers, and social security numbers

Financial Data

Credit card numbers, bank accounts, and payment information

Location Information

Addresses, GPS coordinates, and location-based data

Health Information

Medical record numbers and health-related personal data

Understanding PII Exposure

The dashboard shows:

Exposure Rate: Percentage of sessions that contain PII
Entity Breakdown: Distribution of PII types detected
Location Context: Where PII appears (prompts, responses, tool calls)
Specific Examples: Actual detected values for validation and remediation

⚠️ PII in Production

PII exposure in agent interactions can violate GDPR, CCPA, HIPAA, and other privacy regulations. Use this data to implement proper data masking, tokenization, or access controls before production deployment.

Resource Management

Understand how your agent consumes resources across sessions. Resource management analysis helps prevent cost overruns, performance issues, and runaway executions in production.

Token usage, execution time, and tool call metrics across sessions

Resource Metrics

Token Usage

Average tokens per session
Input vs. output distribution
Token spikes and outliers
Cost projection

Execution Time

Average session duration
Latency patterns
Slow sessions identification
Performance trends

Tool Calls

Calls per session
Tool call sequences
Inefficient patterns
Unused tools

Resource Bounds

Agent Inspector checks whether your agent has appropriate resource bounds configured:

Max Tokens: Ensures token limits are set to prevent runaway costs
Timeout Controls: Verifies execution time limits are in place
Iteration Limits: Checks for controls on multi-turn interactions
Rate Limiting: Identifies missing rate limit protections

Production Readiness Gates

Five mandatory gates that combine security, behavioral, and operational criteria into a clear go/no-go decision for production deployment. Each gate must pass before your agent is considered production-ready.

Production gates showing pass/fail status with specific remediation guidance

The Five Gates

Gate 1: Behavioral Stability

Criterion: Stability Score ≥ 80%

Ensures your agent behaves consistently across sessions. Low stability indicates unpredictable behavior that could cause production issues.

Gate 2: Security Assessment

Criterion: No CRITICAL or HIGH security hazards (or VALIDATED findings)

All critical security findings must be resolved or dismissed with justification. VALIDATED findings from correlation have higher priority.

Gate 3: Supply Chain Integrity

Criterion: Model versions pinned, tools validated

Models must be pinned to specific versions, and all tools must have documented purposes and security posture.

Gate 4: Resource Bounds

Criterion: Token limits, timeouts, and rate limits configured

Prevents runaway costs and resource exhaustion with proper constraints on execution.

Gate 5: Audit & Compliance

Criterion: Critical actions logged, PII handling documented

Ensures compliance requirements are met with proper audit trails and data handling procedures.

Using Gates in CI/CD

Production gates are designed to integrate with your deployment pipeline. Fail the build if gates don't pass, forcing teams to address issues before production deployment. This creates a safety net that catches problems early.

How to Use Risk Assessment

1. Run Sufficient Test Volume

Risk assessment accuracy improves with more data. Different confidence levels require different amounts of data:

20 sessions: Minimum for basic behavioral clustering
30-50 sessions: Ideal for reliable analysis with medium confidence
100+ sessions: Best for complex agents, provides high confidence
200+ sessions: Required for meaningful outlier rate evaluation

Run tests with diverse inputs to capture the full range of agent behavior and potential risks.

2. Test Edge Cases

Include scenarios that stress-test your agent's boundaries:

Malformed or unexpected inputs
Permission edge cases and access control tests
Resource-intensive operations
Error conditions and recovery scenarios
Concurrent or rapid-fire requests

3. Review All Findings

Don't just check the pass/fail status. Review each finding in detail:

Understand root causes of security hazards
Validate PII detections (false positives can occur)
Analyze resource usage patterns for optimization
Document acceptance of warnings and risks

4. Iterate and Re-Test

After addressing findings, run the assessment again to verify fixes:

Confirm critical hazards are resolved
Validate that changes didn't introduce new risks
Ensure behavioral stability improved or remained stable
Verify production gates now pass

5. Generate Reports for Stakeholders

Use the dashboard to create evidence for security reviews and production approvals. The quantitative metrics and specific findings give security teams the data they need to make informed decisions.

Best Practices

Run Early and Often

Don't wait until pre-production. Run risk assessment during development to catch issues when they're cheap to fix.

Integrate with CI/CD

Make risk assessment part of your automated pipeline. Block deployments that fail production gates.

Track Over Time

Compare assessments across versions to ensure changes improve rather than degrade security posture.

Document Decisions

For warnings you choose to accept, document the reasoning and any mitigating controls you've implemented.