figure-it-out
Pass
Audited by Gen Agent Trust Hub on Sep 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill acts as an orchestrator that processes data, artifacts, and self-reports from delegated workers and experimental loops, creating a potential surface for instructions embedded in external artifacts to influence the agent.
- Ingestion points: Phase C (Run the loop) and Phase B (Design the workflow) involve ingesting artifacts from 'delegated work' and 'experiments'.
- Boundary markers: The instructions do not prescribe specific delimiter-based boundary markers for external content, though they recommend independent verification.
- Capability inventory: The skill has the capability to execute scripts for evidence collection (Phase D) and perform file system operations like logging to TSV and committing to a Git repository.
- Sanitization: The skill mitigates this risk by requiring the agent to 'verify by inspecting the artifact, never a self-report' and auditing delegates' artifacts before trusting them.
- [DATA_EXFILTRATION]: The skill encourages the creation and persistence of a detailed 'decision trail' in a TSV format, which may include links to evidence and internal decision logic.
- Evidence: Phase D instructs the agent to log every decision and unit of work into a TSV file and to 'commit it when confidence has to be shown.' While intended for auditability, if the evidence or decision logs include sensitive environment data or internal configurations, committing this trail to a shared repository could lead to unintentional data exposure.
Audit Metadata