recoup-internal-consulting-copy-editor
Pass
Audited by Gen Agent Trust Hub on Oct 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests a draft file and passes its contents to a subagent for review, creating a potential surface for indirect prompt injection if the draft contains adversarial instructions.\n
- Ingestion points: The target draft file path ({ARTICLE_PATH}) is resolved in SKILL.md (Step 2) and its content is read by the subagent in Step 3.\n
- Boundary markers: Absent. The draft content is provided to the subagent without clear delimiters or instructions to ignore embedded commands.\n
- Capability inventory: The main agent has file-writing capabilities (Step 5) and executes a Python script (Step 6). The subagent is correctly configured as readonly in the Claude Code harness, which reduces the risk of unauthorized modifications.\n
- Sanitization: No explicit sanitization or filtering of the draft content is performed before processing.\n- [COMMAND_EXECUTION]: The skill executes a local Python script to perform content scoring and verification.\n
- Evidence: SKILL.md (Step 6) explicitly instructs the agent to run python evals/content/score_run.py to verify the edited piece.
Audit Metadata