runbook-generator
Fail
Audited by Gen Agent Trust Hub on Nov 8, 2026
Risk Level: HIGHINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The
runbook_scaffolder.pyscript accepts JSON service definitions and interpolates user-controlled fields directly into markdown templates, which can be exploited to inject instructions. - Ingestion points: Data is loaded from JSON files or standard input in
scripts/runbook_scaffolder.pyvia theload_service_definitionfunction. - Boundary markers: Absent. The generated runbooks do not use delimiters or explicit warnings to separate template instructions from the interpolated data fields.
- Capability inventory: The skill provides tools for documentation generation, validation, and staleness detection, including
subprocesscalls inscripts/staleness_checker.pyfor git history analysis. - Sanitization: Absent. There is no escaping or validation logic to ensure that strings provided in the JSON (e.g.,
name,dependencies) do not contain malicious markdown or agent directives. - [COMMAND_EXECUTION]: The
staleness_checker.pyscript executes system commands to check the freshness of documentation. - The script uses
subprocess.runto callgit logon configuration files referenced in the runbooks. - It uses the list-based argument format, which is a best practice to prevent shell injection, though it operates on file paths extracted from the runbook content with only basic filtering.
- [EXTERNAL_DOWNLOADS]: The skill generates and contains commands that make network requests to check service health.
- Evidence:
curl -sw "%{{http_code}}" https://$APP_HOST/api/health -o /dev/nullinscripts/runbook_scaffolder.py. - Automated scanners flagged this as a remote code execution pattern. However, analysis shows this is a benign connectivity check that discards the server response body (
-o /dev/null) and only outputs the HTTP status code.
Recommendations
- HIGH: Downloads and executes remote code from: https://$APP_HOST/api/health - DO NOT USE without thorough review
Audit Metadata