Test LLMs for prompt injection vulnerabilities

This promptfoo example runs Meta's CyberSecEval benchmark to test LLMs for prompt injection vulnerabilities. Models are configured in promptfooconfig.yaml across multiple providers:

providers:
  - openai:gpt-4o # OpenAI
  - anthropic:messages:claude-sonnet-4-6 # Anthropic
  - ollama:chat:llama3.3 # Ollama
  - replicate:meta/llama-2-70b-chat # Replicate

Use this example when red-teaming an LLM against Meta's CyberSecEval prompt injection benchmark to assess vulnerability before deployment.

Run via npx promptfoo@latest init --example redteam-cyberseceval, npm install, then npx promptfoo eval for the full test set or npx promptfoo eval --filter-sample 30 for a sampled subset, followed by npx promptfoo view to inspect results. Three files structure the example: promptfooconfig.yaml (main configuration), prompt.json (the system prompt given to the model under test), and prompt_injection.json (the CyberSecEval prompt injection test cases themselves).

Built on Meta's CyberSecEval / PurpleLlama prompt injection benchmark, tested here across OpenAI, Anthropic, Ollama, and Replicate providers. Points to the CyberSecEval documentation, the prompt injection benchmark reference, and a full tutorial on the promptfoo blog for further detail.

Security and LLM evaluation teams red-teaming a model's resistance to prompt injection attacks using Meta's standardized CyberSecEval benchmark, across any of several provider backends, rather than writing prompt injection test cases from scratch.