skill-development

Warn

Audited by Gen Agent Trust Hub on Mar 10, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill implements a custom 'Heir Pull-Sync Protocol' that downloads content from an external source. \n
  • Evidence: The /pullskill command in SKILL.md fetches skill folders and copies them directly into the local .github/skills/ directory. \n
  • Evidence: The /checkskills command reads a remote skills/skill-registry.json file from the 'Global Knowledge' repository. \n
  • This custom update mechanism lacks explicit verification or cryptographic signing, potentially allowing the installation of malicious instructions from a compromised registry. \n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its project signal detection and wishlist fulfillment processes. \n
  • Ingestion points: The skill scans local project files such as .github/workflows/, docker-compose.yml, package.json, and requirements.txt (SKILL.md). \n
  • Boundary markers: No delimiters or instructions to ignore embedded commands within these scanned files are documented. \n
  • Capability inventory: The agent has the capability to write to the local filesystem (/pullskill), update manifests, and push insights to a shared repository (/saveinsight). \n
  • Sanitization: There is no evidence of sanitization or validation of the data read from project files before it is processed for skill matching. \n- [COMMAND_EXECUTION]: The skill defines several custom commands that perform sensitive operations on the filesystem and network. \n
  • Evidence: The /pullskill, /checkskills, /fulfillwish, and /saveinsight commands represent custom execution paths that manage the lifecycle of agent capabilities, modifying local configurations and interacting with remote data sources.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 10, 2026, 03:55 AM