Crucible
crucible-security.github.io
- Free plan
- Linux
- Self-hosted
What Crucible is
Crucible is a self-hosted security tool for stress-testing AI agents against attacks such as prompt injection, goal hijacking, and tool misuse before production. It evaluates tool calls, side effects, and outcomes, not just generated text. The site lists 11 modules and 153 attack vectors, covering eight of ten OWASP Agentic risks, with two in active development. Crucible says it runs 90 attack payloads against an agent in under 60 seconds and returns an A–F graded report with severity, OWASP mapping, and reproduction steps. It supports LangChain, CrewAI, AutoGen, raw MCP endpoints, and agents with an HTTP or MCP interface. The CLI installs with pip and supports Python 3.10 or later; the site says starting requires no Docker, cloud signup, or API key. Local scans require no signup and have no telemetry by default. The Apache 2.0 core engine and test modules are free, with no feature gating on security tests. Community costs nothing.
Who it suits
Crucible suits individual builders and open-source projects using the Community plan, as well as teams preparing agents for production. It is aimed at users who want to run security checks locally or in CI.
Good to know, and what to watch
- Lists 153 attack vectors across 11 modules
- Produces graded reports with severity and reproduction steps
- Supports several agent frameworks and HTTP or MCP interfaces
- GitHub Actions can fail builds below a configured score
- Local scanning has no telemetry by default
- Two of ten OWASP Agentic risks are in active development
- The CLI requires Python 3.10 or later
Plans
Every plan and priceEnterprise
Team
Community
Free plan
The verdict
Crucible offers agent-focused security testing with local and CI workflows.
What the maker publishes
- Attack coverage
- The site lists 11 modules and 153 attack vectors, with 8 of 10 OWASP Agentic risks covered and two in active development. source
- CI/CD
- The GitHub Actions integration can fail a build when an agent's security score drops below a configured threshold. source
- Install requirements
- The CLI installs with pip, supports Python 3.10+, and the site says no Docker, cloud signup, or API key is needed to start. source
- Integrations
- It supports LangChain, CrewAI, AutoGen, raw MCP endpoints, and agents with an HTTP or MCP interface. source
- Intended users
- The Community plan is for individual builders and open-source projects, and Team is for teams shipping agents to production. source
- License
- The core engine and test modules are described as Apache 2.0 and free, with no feature gating on security tests. source
- Maker
- The team page names Khalid Saifullah Sayyed as Founder & Core and says the team was built in 2026. source
- Privacy
- The site says local scanning requires no signup and has no telemetry by default. source
- Purpose
- Crucible stress-tests AI agents against attack vectors including prompt injection, goal hijacking, and tool misuse before production. source
- Reports
- Scans produce an A–F graded report with severity, OWASP mapping, and reproduction steps. source
- Scan speed
- Crucible says it runs 90 attack payloads against an agent in under 60 seconds. source
- Support
- The Community plan lists Discord support, while Team lists priority support and Enterprise lists dedicated security engineering support. source
- Testing approach
- Crucible evaluates actual agent tool calls, side effects, and outcomes rather than only generated text. source
Checked against the maker's pages on 5 Oct 2026.