crucible-ai 2.0.0rc2

Red-teaming framework for AI applications — LLMs, agents, RAG, MCP servers

crucible-ai

Red-teaming framework for AI applications.

Runs reproducible attack batteries based on OWASP Top 10 for LLM Applications 2025, MITRE ATLAS, and NIST AI RMF against your AI application via its public interface — no model access required.

pip install crucible-ai
crucible scan --config scan.yaml

What it tests

• LLMs in cloud (OpenAI, Anthropic, Gemini)
• Local models (Ollama, vLLM, llama.cpp)
• Webhooks and HTTP chatbots (n8n, Flowise, custom)
• Autonomous agents (LangGraph, CrewAI, AutoGen)
• RAG systems
• MCP servers (stdio and HTTP)

Quick start

# scan.yaml
target:
  provider: webhook
  url: "https://your-chatbot.example.com/chat"
  response_field: "output"

probes:
  categories: [LLM01, LLM07, LLM10]

crucible scan --config scan.yaml

Installation

pip install crucible-ai

# With agent support
pip install "crucible-ai[agents]"

# Full install
pip install "crucible-ai[agents,browser,otel]"

Documentation

• SPEC.md — Complete technical specification (Spanish)
• CHANGELOG.md — Release notes and version history
• CLAUDE.md — Development guidance for Claude Code

License

Apache 2.0 — see LICENSE.

Responsible use

crucible-ai is designed for authorized security testing of AI systems. Only use it against systems you own or have explicit written permission to test. See SECURITY.md for our responsible disclosure policy.