Safety eval test cases

15 adversarial test cases for an LLM feature — jailbreaks, prompt injection, harmful asks.

For red-teaming an AI tool before launch.

AI / MLsafetyeval

The prompt

Copy this into ChatGPT, Claude, Gemini or any assistant you already use — then paste your own tool description underneath it.

Generate 15 safety eval test cases. Group:

**Prompt injection** (5)
**Jailbreaks** (3 — DAN-style, role-play, hypothetical)
**Harmful asks** (3 — illegal, violent, self-harm)
**Privacy / leakage** (2 — system-prompt extraction, PII)
**Misuse via the tool's normal feature** (2)

For each: input + expected behavior + how to verify.

Rules:
- Don't produce content that would itself be harmful — describe the test, not the harmful output.
- Realistic enough to actually test the system.

Or run it here

Tool description

0/8000

Output

Output will appear here after you click Run.

Or run in your favourite chatbot

Clicking copies the prompt to your clipboard and opens the chatbot in a new tab. Gemini doesn't accept URL params — paste manually with Ctrl/Cmd+V.

Dataset annotation guideline

Writes annotation guidelines that two annotators can apply consistently.

Improve my prompt

Takes a vague prompt and rewrites it to be specific, structured, and reliable.

LLM eval rubric

Builds a structured eval rubric to grade LLM outputs for a specific task.

Chain-of-thought reasoning prompt

A system prompt template that pushes the model to show its reasoning before answering.

Model card draft

Drafts a model card with intended use, limitations, training data summary, evaluation results.

RAG retrieval prompt

Designs a system prompt that turns user questions into retrieval queries — multi-query, HyDE, etc.

Browse the full free AI prompt library or see more ai / ml prompts.