pr-gfm-validator

Warn

Audited by Gen Agent Trust Hub on Oct 8, 2026

Risk Level: MEDIUMPERSISTENCEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONCREDENTIALS_UNSAFE
Full Analysis
  • [PERSISTENCE]: The skill contains 'Self-Evolving' instructions in SKILL.md that explicitly direct the agent to 'fix this file immediately' if issues occur. Combined with the 'Post-Execution Reflection' which tells the agent to locate and edit the SKILL.md file, this creates a persistence mechanism. If an agent is subverted via prompt injection, it could modify its own source code to include malicious behavior that persists across all future sessions.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted pull request descriptions and body content. There is a high risk that malicious instructions embedded in a PR body could influence the 'Self-Evolving' logic to modify the skill's instructions.
  • Ingestion points: PR descriptions and body content processed during gh pr create or link validation workflows.
  • Boundary markers: None. The skill does not use delimiters or instructions to ignore potential commands within the PR body.
  • Capability inventory: Access to Bash tool (executing gh and git), and implied file-write access to its own SKILL.md file.
  • Sanitization: No sanitization or validation is performed on the PR body content before it is processed or used in the evolution loop.
  • [COMMAND_EXECUTION]: The skill uses Bash to execute gh (GitHub CLI) and git commands. While these are used for intended repository management, the interpolation of CONVERTED_BODY (derived from untrusted PR text) into the gh pr create command presents a command injection surface if the agent does not properly handle shell escaping.
  • [CREDENTIALS_UNSAFE]: The skill's documentation (Example 3) explicitly references sensitive credential files (.env.clickhouse). While used as an example for link conversion, referencing specific environment file paths encourages the agent to interact with or expose locations containing sensitive secrets.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 8, 2026, 02:12 AM