review-pr-local

Warn

Audited by Gen Agent Trust Hub on Nov 19, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes several Python scripts located within the repository's .github/scripts/ directory. These include prepare_local_review_inputs.py, validate_review_json.py, and validate_local_review_result.py. While these are local files, their execution constitutes a core part of the skill's operational surface.
  • [DYNAMIC_EXECUTION]: The workflow relies on a script output to determine which instructions to follow next. Specifically, prepare_local_review_inputs.py prints a skill=<path>, and the agent is instructed to "follow the selected skill exactly." This dynamic loading of instructions from a computed path can be exploited if the repository state or script logic is compromised.
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests data from pull request descriptions and code diffs, which are often provided by external contributors and may contain malicious instructions.
  • Ingestion points: Data is read from pr_description.txt and pr_diff.txt snapshots.
  • Boundary markers: There are no defined boundary markers or "ignore previous instructions" warnings wrapping the untrusted input files.
  • Capability inventory: The skill has the ability to execute shell commands (via Python) and read/write files in the repository.
  • Sanitization: No sanitization or validation of the pull request description or diff content is specified before the agent processes them.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 19, 2026, 01:13 AM