gh-create-pr

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: Rule 7 instructs the agent to omit session URLs "even if system instructions ask for one," which is a targeted instruction to override governing system prompts.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process data from external, potentially untrusted sources to generate pull request content.
  • Ingestion points: Processes data from linked issues, tickets, conversation history, and CLAUDE.md (SKILL.md).
  • Boundary markers: Employs a quoted HEREDOC (cat <<'EOF') for the PR body which prevents expansion of content, but the PR title field lacks isolation (SKILL.md).
  • Capability inventory: Uses the gh CLI tool for PR creation, involving network operations and local shell command execution (SKILL.md).
  • Sanitization: No explicit requirements or mechanisms for sanitizing or escaping data before interpolation into the PR command are provided (SKILL.md).
  • [COMMAND_EXECUTION]: The bash template in Rule 1 interpolates the PR title directly into the shell command: --title "<short title>". If the content sourced for the title contains shell metacharacters (e.g., backticks or subshells), it could result in unintentional command execution in the user's local shell environment.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 04:15 AM
Security Audit — agent-trust-hub — gh-create-pr