workflow-score-to-target
Pass
Audited by Gen Agent Trust Hub on Sep 28, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines a workflow that ingests external data (renders, recordings, and measurements) and stores the resulting analysis in project files or memory to guide future agent actions. This creates a potential surface where malicious instructions embedded within the analyzed artifacts could influence the agent's judging logic or its persistent project context.\n
- Ingestion points: The agent scores work based on "renders, recordings or measurements of the actual thing" as described in Section 3 of SKILL.md.\n
- Boundary markers: The workflow utilizes a fresh subagent with a dedicated critic prompt (templates/critic-prompt.md) to judge work, providing a layer of isolation for the scoring task.\n
- Capability inventory: The skill is instructed to record targets in project instructions or memory (Section 1) and save detailed scorecards to the file system, such as in "qa//README.md" (Section 5).\n
- Sanitization: There are no explicit instructions to sanitize the content of the artifacts being judged or to ignore potential command instructions embedded within them.
Audit Metadata