crucible-security/Crucible Security Agent Scan

Security scan for AI agents. Audits prompt injection, jailbreaks, and tool misuse in GitHub workflows.

View on GitHub

Trust Signals

Scorecard Score
not yet scored
Maintenance Recency
Stale
License
None
namedescriptionrequireddefault
targetURL of the AI agent endpoint to scan (e.g. https://my-agent.example.com/api/chat). yes
format_presetProvider format preset. One of: openai, anthropic, ollama, lmstudio, gemini. noopenai
modelModel name to send in the request body (e.g. gpt-4o, claude-3-5-sonnet, llama3.2). no""
headersExtra HTTP headers as a JSON object string (e.g. '{"Authorization": "Bearer sk-..."}'). Use GitHub Secrets for sensitive values. no{}
modulesComma-separated list of modules to run. Leave empty to run all modules. Available: prompt_injection, goal_hijacking, jailbreaks, mcp_security, tool_injection, enterprise_graph, memory_poisoning, advanced_orchestration, infrastructure_escalation, agentic_hijacking, browser_agent, hallucination, toxicity, inter_agent_trust no""
fail_on_gradeFail the workflow if the scan grade is BELOW this value. Use A, B, C, D, F. Default is F (never fails on grade alone). Example: set to C to fail on C, D, or F grades. noF
concurrencyMaximum concurrent attack requests.no10
timeoutHTTP request timeout in seconds.no30
versionVersion of crucible-security to install. Defaults to latest. no""
output_formatFormat of the uploaded scan report artifact. One of: json, sarif. nojson
upload_artifactSet to 'true' to upload the scan report as a workflow artifact. notrue
artifact_nameName of the uploaded artifact. nocrucible-scan-report
skip_preflightSet to 'true' to skip the endpoint preflight check. nofalse
namedescription
gradeSecurity grade (A, B, C, D, or F).
scoreSecurity score (0.0 – 100.0).
failed_countNumber of failing attack findings.
total_countTotal number of attack findings run.
report_pathAbsolute path to the generated report file.