agent.injectionInjection & data-flow review
Traces untrusted input from sources to dangerous sinks (SQL/NoSQL, shell, outbound HTTP, HTML, filesystem, templates, deserializers) and reports only reachable, unsanitised paths. (agentic, uses Claude)
- Engine
- Agent
- Targets
- Codebase
- Areas
- Security
- Runs
- Opt-in: only when selected with --only
- Network
- Contacts services beyond the target
Run only this auditor
bunx openaudit . --only agent.injectionRuns a playbook with the Claude Agent SDK. Needs ANTHROPIC_API_KEY or a Claude Code login; host agents can run the same playbook over MCP without a key.
Playbook
Traces untrusted input from sources to dangerous sinks (SQL/NoSQL, shell, outbound HTTP, HTML, filesystem, templates, deserializers) and reports only reachable, unsanitised paths.
Host agents fetch it with get_playbook over MCP. How playbooks work