polish-batch
Pass
Audited by Gen Agent Trust Hub on Oct 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes 'Wrong -> Right' strings from a markdown punch-list file and applies them as code changes. This creates a surface for indirect prompt injection if the source content is untrusted. Ingestion points: Data is read from the -punchlist.md file during the dispatch and verify phases. Boundary markers: Instructions specify 'Fix exactly these listed items... and nothing else,' providing a control against scope creep. Capability inventory: The skill uses git tools, browser automation, and a 'coding CLI' for file modifications. Sanitization: There is no explicit sanitization for the input strings before they are used in code edits.
- [COMMAND_EXECUTION]: The skill performs shell command execution for core tasks, including git operations and invoking a coding CLI. The evaluation documentation also references the use of local bash and python scripts for test execution.
Audit Metadata