find-skills

Fail

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: HIGHEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPERSISTENCEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill downloads a 'bootstrap' kit from a Tencent Cloud COS bucket (https://skillhub-1388575217.cos.ap-guangzhou.myqcloud.com/install/latest.tar.gz) and individual skills from the domain lightmake.site. While Tencent Cloud is a known infrastructure provider, lightmake.site is an unverified third-party domain used for search and primary downloads.
  • [REMOTE_CODE_EXECUTION]: The bootstrap command downloads Python scripts (skills_store_cli.py, skills_upgrade.py) and installs them directly into the user's home directory (~/.skillhub). These scripts are subsequently executed by the agent to perform management tasks, constituting a remote code execution pattern where untrusted external code is fetched and run.
  • [COMMAND_EXECUTION]: The script uses subprocess.run and subprocess.Popen to interact with an openclaw binary and to execute the newly installed skillhub CLI. It programmatically modifies configuration settings and restarts the openclaw gateway service.
  • [PERSISTENCE]: The installation process writes wrapper scripts (skillhub, oc-skills) to common binary directories (~/.local/bin on Linux/macOS or AppData on Windows) and applies executable permissions (chmod 0o755). This ensures that the downloaded tools remain available and executable across different user sessions.
  • [INDIRECT_PROMPT_INJECTION]: The search command fetches skill names and descriptions from the external API at lightmake.site. This metadata is ingested and presented to the agent, creating a surface for potential instructions embedded in external data to influence the agent's behavior during the installation workflow.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Sep 15, 2026, 01:59 AM
Security Audit — agent-trust-hub — find-skills