# LLM Safety Evaluation and Policy Test Suites
Status: public
Confidence: medium (0.725) (verified)
Last verified: 2026-06-02
Generation: ai_structured


## TL;DR

LLM safety test suites make policy behavior measurable before a model, prompt, or agent workflow reaches production.

## Core Explanation

Safety evaluation is not one prompt that asks whether a system is safe. It is a set of policy-grounded test cases, adversarial probes, monitoring signals, and review workflows tied to the application's risk profile.

Agents should report which policy categories were tested, which evaluators were used, what threshold failed or passed, and which examples require human review. Safety regressions need specific evidence rather than a generic "guardrail passed" claim.

## Source-Mapped Facts

- Azure AI Foundry documentation describes risk and safety evaluators for assessing AI application outputs. ([source](https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/evaluation-evaluators/risk-safety-evaluators))
- OpenAI safety best-practices documentation describes safety work as including evaluations and monitoring. ([source](https://platform.openai.com/docs/guides/safety-best-practices))
- Promptfoo red-team documentation describes red teaming as testing LLM applications for vulnerabilities. ([source](https://www.promptfoo.dev/docs/red-team/))

## Further Reading

- [Azure AI Foundry Risk and Safety Evaluators](https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/evaluation-evaluators/risk-safety-evaluators)
- [OpenAI Safety Best Practices](https://platform.openai.com/docs/guides/safety-best-practices)
- [Promptfoo Red Teaming](https://www.promptfoo.dev/docs/red-team/)