AI工作流

Install

pip install crucible-security

Quick Start

🆕 New to AI security? Read our Beginner's Getting Started Guide.

crucible init --target https://my-agent.com/api/chat
crucible scan --target https://my-agent.com/api/chat
crucible report crucible-report.json

One command. 90 attacks. Beautiful report.

Examples

We provide several example scripts in the examples/ directory to help you get started:

All examples use respx to mock HTTP calls so they pass CI without a live server.

Running the LangChain Example:

python examples/test_langchain_agent.py

Running the OpenAI Assistant Example:

python examples/test_openai_assistant.py

Generate config

crucible init --target URL --provider openai --key sk-xxx

CLI Reference

```bash

Modules

CI/CD Integration

Add to your CI/CD in 3 lines:

```yaml

.github/workflows/security.yml

- uses: actions/checkout@v4 - run: pip install crucible-security - run: crucible scan --target ${{ secrets.AGENT_URL }} --fail-on CRITICAL ```

FAQ

Does Crucible send my agent data to your servers? No. Crucible is a local CLI. Payloads go directly from your machine to your agent. Nothing passes through Crucible infrastructure. Zero data retention. Fully air-gappable.

Which agent frameworks does Crucible support? Any agent that accepts HTTP requests — LangChain, AutoGen, CrewAI, OpenAI Assistants, Bedrock, custom FastAPI agents.

How long does a full scan take? Under 60 seconds for 90 attacks using async parallel execution.

Can I add custom attack vectors? Yes. See CONTRIBUTING.md for how to submit new attack modules via PR.

Is this safe to run against production? Run against staging environments, not production. Crucible sends adversarial payloads that may cause unexpected behavior.

What does Grade F mean? Your agent complied with most attacks. It is vulnerable to prompt injection, jailbreaks, or goal hijacking. Review Critical findings first.

Why is the module called goal_hijacking if goal hijacking is an impact, not an attack? Crucible modules are named by the security impact they surface, not the attack vector. The underlying attack vector for most modules is prompt injection delivered in specialised forms. This naming convention helps security engineers quickly identify which risks each module addresses (e.g., searching for "goal hijacking" finds the right module immediately). See docs/owasp_mapping.md for the full attack vector → impact mapping.

Questions not answered here? Join our Discord or email crucible.sec@gmail.com