Skip to main content

​
What is a Harness?

A harness is a collection of tests bundled together for a specific purpose. When you run an evaluation, you select one or more harnesses. Each harness produces a score and detailed results for everything it tests. Think of a harness like a test suite in traditional software testing. A unit test suite tests individual functions. An integration test suite tests components working together. A security test suite tests for vulnerabilities. Each suite has a different purpose and contains different tests—but they all use the same testing infrastructure. Vijil harnesses work the same way. The reliability harness tests for correctness, consistency, and robustness. The owasp_llm_top_10 harness tests for the vulnerabilities in the OWASP LLM Top 10. The trust_score harness runs everything and produces a comprehensive Trust Score. Different purposes, same evaluation infrastructure.

​
Types of Harnesses

​
Dimension Harnesses

Each trust dimension has its own harness:
HarnessWhat It Tests
reliabilityCorrectness, consistency, robustness
securityConfidentiality, integrity, availability
safetyContainment, compliance, transparency
These harnesses map directly to the Trust Score dimensions. Run all three to get a complete Trust Score, or run individual dimensions to focus on specific concerns.

​
Compliance Harnesses

Compliance harnesses test against external standards:
HarnessWhat It Tests
owasp_llm_top_10OWASP LLM Top 10 vulnerabilities
nist_ai_rmfNIST AI Risk Management Framework controls
gdprGDPR-relevant privacy and data handling
These harnesses are designed to produce evidence for audits and compliance reviews. Results map to specific controls or requirements in each standard.

​
Benchmark Harnesses

Benchmark harnesses test against established AI evaluation benchmarks:
HarnessWhat It Tests
openllm_v2Tasks from the Open LLM Leaderboard v2
strongrejectStrongREJECT jailbreak resistance benchmark
cybersecevalMeta’s CyberSecEval security benchmark
Benchmark harnesses let you compare your agent against published results and track performance on standardized tasks.

​
The Trust Score Harness

The trust_score harness is a meta-harness that includes all dimension harnesses plus performance benchmarks. It produces the complete Vijil Trust Score—the single number that captures overall trustworthiness. 
# Run a complete Trust Score evaluation
from vijil import Vijil

client = Vijil()
evaluation = client.evaluations.create(
    agent_id="your-agent-id",
    harnesses=["trust_score"]
)

​
Custom Harnesses

Standard harnesses test for general-purpose trust. But your agent has specific characteristics—a particular system prompt, defined capabilities, a knowledge base, restricted topics. Custom harnesses let you test for these specifics. A custom harness can:
  • Test your agent’s adherence to its system prompt
  • Verify it stays within its defined capabilities
  • Check that it correctly uses (or refuses to use) specific tools
  • Probe for leakage of information from your knowledge base
  • Test compliance with your organization’s specific policies
Custom harnesses use the same scenarios, probes, and detectors as standard harnesses, but configured for your context.

Building Custom Harnesses

Learn how to create harnesses tailored to your agent

​
Harness Results

When a harness completes, you get:
  • Overall score: A 0-100 score for everything the harness tests
  • Scenario breakdown: Scores for each scenario within the harness
  • Probe results: Pass/fail for individual test cases
  • Failure details: Exactly which probes failed and why
Results are hierarchical. You can see the top-level score, drill into scenarios to find problem areas, and examine individual probes to understand specific failures.

​
Next Steps

Scenario

How probes are grouped within harnesses

Standard Harnesses

Reference for available harnesses

Run an Evaluation

Try harnesses on your agent

Custom Harnesses

Build harnesses for your use case
Last modified on March 19, 2026