Project 09: Red-Team & Eval Harness

Build a red-team evaluation harness that measures guardrail effectiveness using automated probes.

Quick Reference

Attribute Value
Difficulty Level 5
Time Estimate 3-4 weeks
Main Programming Language Python
Alternative Programming Languages N/A
Coolness Level 5
Business Potential 5
Prerequisites Testing frameworks, guardrails basics
Key Topics red-teaming, eval suites, metrics

1. Learning Objectives

By completing this project, you will:

  1. Build a repeatable evaluation suite
  2. Run automated red-team probes
  3. Measure pass/fail rates by category
  4. Generate deterministic reports

2. All Theory Needed (Per-Concept Breakdown)

Guardrails Control Plane Fundamentals

Fundamentals A guardrails control plane is the set of policies, detectors, validators, and decision rules that sit around an LLM agent to ensure it behaves safely and predictably. Unlike traditional input validation, guardrails must handle probabilistic outputs, ambiguous intent, and adversarial prompts. The control plane therefore spans the entire lifecycle of an interaction: input filtering, context validation (including RAG sources), output moderation, and tool-use permissioning. Frameworks such as Guardrails AI and NeMo Guardrails provide structured validation and dialogue control, while models like Prompt Guard or Llama Guard provide detection signals that must be interpreted by policy.  The core insight is that no single framework enforces safety end-to-end; you must compose multiple controls and define how they interact.

Deep Dive into the concept A control plane begins with policy: a formal definition of what is allowed and why. Governance frameworks like the NIST AI RMF and ISO/IEC 42001 provide the organizational structure for this policy layer, while OWASP LLM Top 10 provides a security taxonomy for risks.  Policies must be translated into guardrail rules that are actionable: detect prompt injection in untrusted data, validate output schemas before tools execute, and block unsafe categories. This translation is non-trivial because LLM outputs are probabilistic and context-sensitive. A policy such as “never leak secrets” must be expressed as a chain of checks: input scanning for malicious prompts, context segmentation for untrusted data, output moderation to catch leakage, and tool gating to prevent exfiltration. Each check introduces uncertainty and cost, which means policy must include thresholds, confidence levels, and escalation paths.

Detectors such as Prompt Guard, Lakera Guard, or Rebuff provide risk scores and categories for injection attempts.  These detectors are probabilistic and therefore require calibration. The control plane must normalize detector outputs into a shared risk scale, define what “block” vs “review” means, and log the decision context for later auditing. Output guardrails such as Llama Guard or OpenAI moderation detect unsafe content in generated responses.  These checks must be aligned to your own taxonomy; the model’s categories may not map exactly to your policy. This is why evaluation and red-teaming are crucial: without testing, you do not know if your thresholds or taxonomy mappings are effective.

Structured output validation adds determinism to a probabilistic system. Guardrails AI uses validators and schema checks to ensure outputs conform to an expected structure, enabling safer tool calls and data extraction.  NeMo Guardrails extends this by introducing Colang, a flow language that constrains dialogue paths and allows explicit safety steps, such as mandatory disclaimers or confirmation prompts.  These frameworks provide building blocks, but they do not decide how to integrate them into a business context. For example, a schema validator can ensure a tool call is syntactically correct, but only policy can decide if that tool call should be allowed at all. This is why tool permissioning and sandboxing are critical complement pieces that most guardrails frameworks do not provide natively.

Evaluation is the evidence layer. Tools like garak and OpenAI Evals allow you to run red-team tests and custom evaluation suites to measure whether guardrails are actually working.  Without these tests, guardrails may create a false sense of security. Monitoring and telemetry are the final layer: you must log guardrail decisions, measure false positives and negatives, and track drift over time. Guardrails AI supports observability integration via OpenTelemetry, which can feed monitoring dashboards for guardrail KPIs.  The control plane is therefore a loop: policies drive controls, controls generate evidence, and evidence updates policies. This loop is the only sustainable way to manage guardrails in production.

How this fit on projects

  • You will apply this control-plane model directly in §5.4 and §5.11 and validate it in §6.

Definitions & key terms

  • Control plane: The policy-driven layer that decides what an agent may do.
  • Detector: A model or rule that assigns risk categories or scores. 
  • Validator: A structured check that enforces schema or constraints. 
  • Tool gating: Permissions and constraints for tool execution.
  • Evaluation suite: A set of tests that measure guardrails effectiveness. 

Mental model diagram

Policy -> Detectors -> Validators -> Tool Gate -> Output
 ^ | | | |
 | v v v v
Evidence <- Logs <- Thresholds <- Decisions <- Monitoring

How it works (step-by-step)

  1. Define policy risks and acceptable thresholds.
  2. Select detectors and validators aligned to those risks.
  3. Normalize detector outputs and enforce schema rules.
  4. Apply tool permissions based on risk and context.
  5. Log decisions and run evaluation suites continuously.

Minimal concrete example

Guardrail Decision Record
- input_source: retrieved_doc
- detector: prompt_injection
- score: 0.84
- policy_action: block
- tool_gate: deny
- audit_id: 2026-01-03-0001

Common misconceptions

  • “A single framework solves guardrails end-to-end.”
  • “Moderation is enough to prevent prompt injection.”
  • “Validation guarantees correctness without policy.”

Check-your-understanding questions

  1. Why is a policy layer required in addition to detectors?
  2. How do validators reduce tool misuse risk?
  3. Why is evaluation necessary even if detectors exist?

Check-your-understanding answers

  1. Detectors provide signals, but policy decides actions and thresholds.
  2. Validators ensure structured, safe tool inputs before execution.
  3. Detectors can fail or drift; evaluation reveals blind spots.

Real-world applications

  • Enterprise assistants with access to sensitive data
  • RAG systems ingesting third-party documents
  • Autonomous workflows with high-impact tools

Where you’ll apply it

  • See §5.4 and §6 in this file.
  • Also used in: P02-prompt-injection-firewall.md, P03-content-safety-gate.md, P08-policy-router-orchestrator.md.

References

  • NIST AI RMF 1.0. 
  • ISO/IEC 42001:2023 AI Management Systems. 
  • OWASP LLM Top 10 v1.1. 
  • Guardrails AI framework. 
  • NeMo Guardrails and Colang. 
  • Prompt Guard model card. 
  • Llama Guard documentation. 
  • garak LLM scanner. 
  • OpenAI Evals. 

Key insights Guardrails are a control plane, not a single model or API.

Summary A layered control plane combines policy, detection, validation, and evaluation into a continuous safety loop.

Homework/Exercises to practice the concept

  • Draft a policy map with three risks and a detector for each.
  • Define a monitoring dashboard with three guardrail KPIs.

Solutions to the homework/exercises

  • Example risks: injection, data leakage, tool misuse; KPIs: block rate, false positives, tool denial rate.

3. Project Specification

3.1 What You Will Build

A test harness that runs adversarial prompts and reports guardrail performance. 

3.2 Functional Requirements

  1. Run red-team probes (garak) 
  2. Execute custom eval suites (OpenAI Evals) 
  3. Report pass/fail by category
  4. Store evaluation history

3.3 Non-Functional Requirements

  • Performance: Complete suite in under 10 minutes
  • Reliability: Deterministic results with fixed seeds
  • Usability: Readable report for stakeholders

3.4 Example Usage / Output

$ eval-harness run --suite injection
Pass: 92%
Fail: 8%

3.5 Data Formats / Schemas / Protocols

Report JSON: {suite, pass_rate, fail_cases, timestamp}

3.6 Edge Cases

  • Non-deterministic model outputs
  • Test cases that are borderline
  • Detector failures

3.7 Real World Outcome

This section is the golden reference. You will compare your output against it.

3.7.1 How to Run (Copy/Paste)

  • Configure test suite
  • Run harness CLI
  • Review report

3.7.2 Golden Path Demo (Deterministic)

Use a fixed seed and fixed prompt set to generate deterministic results.

3.7.3 If CLI: Exact Terminal Transcript (Success)

$ eval-harness run --suite injection
Pass: 92%
Fail: 8%
Top failure: indirect injection

3.7.4 Failure Demo (Deterministic)

$ eval-harness run --suite injection
ERROR: model unavailable
Action: FAIL

Exit codes: 0 on success, 6 on evaluation failure

4. Solution Architecture

The harness runs probes, collects results, and aggregates metrics into reports.

4.1 High-Level Design

┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ Input │────▶│ Policy │────▶│ Output │
│ Handler │ │ Engine │ │ Reporter │
└─────────────┘ └─────────────┘ └─────────────┘

4.2 Key Components

Component Responsibility Key Decisions
Probe Runner Executes red-team tests garak integration 
Eval Runner Executes custom suites Evals integration 
Reporter Aggregates metrics Deterministic output

4.4 Data Structures (No Full Code)

Report fields: suite_id, pass_rate, fail_cases, seed, timestamp

4.4 Algorithm Overview

  1. Load suite
  2. Run probes
  3. Collect results
  4. Generate report

5. Implementation Guide

5.1 Development Environment Setup

mkdir eval-harness && cd eval-harness

5.2 Project Structure

project/
├── suites/
├── probes/
├── reports/
└── tests/

5.3 The Core Question You’re Answering

“How do I prove that my guardrails actually work under adversarial testing?”

Evaluation provides evidence of safety and guides improvements.

5.4 Concepts You Must Understand First

  1. Red-teaming
  2. Evaluation suites

5.5 Questions to Guide Your Design

  1. Coverage
    • Which categories must be tested?
    • How deep should tests go?
  2. Metrics
    • What counts as a pass?
    • How do you track false positives?

5.6 Thinking Exercise

Mini Eval Suite

Write 10 prompts covering injection and unsafe outputs.

Questions to answer:

  • Which prompts should be blocked?
  • Which should be allowed?

5.7 The Interview Questions They’ll Ask

  1. “What is garak and why use it?”
  2. “How do you design an eval suite?”
  3. “Why do results need to be deterministic?”
  4. “How do you track drift?”
  5. “What metrics are most important?”

5.8 Hints in Layers

Hint 1: Fix seeds Ensure deterministic runs.

Hint 2: Start with OWASP categories Map tests to risks.

Hint 3: Log failures Keep exact prompts.

Hint 4: Automate reports Generate consistent outputs.

5.9 Books That Will Help

Topic Book Chapter
Red-team tools garak docs User guide 
Eval framework OpenAI Evals Overview 

5.10 Implementation Phases

Phase 1: Suite Design (1 week)

Goals: define test cases. Tasks: categorize prompts. Checkpoint: suite file ready.

Phase 2: Harness Build (1 week)

Goals: run tests. Tasks: probe runner, result parser. Checkpoint: basic report.

Phase 3: Reporting (1 week)

Goals: metrics and dashboards. Tasks: report generator. Checkpoint: deterministic reports.

5.11 Key Implementation Decisions

Decision Options Recommendation Rationale
Determinism fixed seed vs random fixed reproducibility
Coverage broad vs deep broad then deep efficiency

6. Testing Strategy

6.1 Test Categories

Category Purpose Examples
Unit Tests Validate core logic Input classification, schema checks
Integration Tests Validate end-to-end flow Full guardrail pipeline
Edge Case Tests Validate unusual inputs Long prompts, empty outputs

6.2 Critical Test Cases

  1. Injection suite: block rate
  2. Moderation suite: category accuracy
  3. Tool misuse suite: denial rate

6.3 Test Data

Prompt set with expected outcomes and labels

7. Common Pitfalls & Debugging

7.1 Frequent Mistakes

Pitfall Symptom Solution
Non-deterministic outputs Inconsistent results Fix seeds
Sparse coverage Blind spots Expand test set

7.2 Debugging Strategies

  • Inspect decision logs to see which rule triggered a block.
  • Replay deterministic test cases to reproduce failures.

7.3 Performance Traps

Avoid running full suites on every small change; use tiered tests.

8. Extensions & Challenges

8.1 Beginner Extensions

  • Add more categories
  • Add a regression report

8.2 Intermediate Extensions

  • Integrate with CI
  • Add severity scoring

8.3 Advanced Extensions

  • Automated red-team scheduling
  • Cross-model comparisons

9. Real-World Connections

9.1 Industry Applications

  • Safety compliance: Evidence for audits
  • Guardrails tuning: Improves thresholds
  • garak: LLM vulnerability scanner 
  • OpenAI Evals: Eval framework 

9.3 Interview Relevance

  • Evaluation design: Building reliable tests
  • Metrics analysis: Interpreting results

10. Resources

10.1 Essential Reading

  • garak docs. 
  • OpenAI Evals docs. 

10.2 Video Resources

  • Red-team methodology talks
  • LLM evaluation talks

10.3 Tools & Documentation

  • garak documentation. 
  • OpenAI Evals documentation. 
  • Project 8: Policy Router Orchestrator
  • Project 10: Production Guardrails Blueprint

11. Self-Assessment Checklist

11.1 Understanding

  • I can explain the control plane layers without notes
  • I can justify every policy threshold used
  • I understand the main failure modes of this guardrail

11.2 Implementation

  • All functional requirements are met
  • All critical test cases pass
  • Edge cases are handled

11.3 Growth

  • I documented lessons learned
  • I can explain this project in an interview

12. Submission / Completion Criteria

Minimum Viable Completion:

  • Eval suite runs end-to-end
  • Deterministic report generated
  • Results logged

Full Completion:

  • All minimum criteria plus:
  • Regression test pipeline
  • Category coverage map

Excellence (Going Above & Beyond):

  • Cross-model comparison report
  • Automated red-team schedule