skill-picker

Pass

Audited by Gen Agent Trust Hub on Nov 5, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands such as git ls-files, find, grep, mkdir, and ln -s for project inventory and symlink management during installation. It also utilizes npx skills and uvx library-skills to interact with external skill registries.
  • [EXTERNAL_DOWNLOADS]: The skill fetches remote instructions and scripts from GitHub and public registries using npx, uvx, git clone, and the GitHub API. These downloads are directed toward third-party repositories and the author's own skill catalog.
  • [INDIRECT_PROMPT_INJECTION]: The skill inventories the project by reading untrusted data sources including manifests (package.json, pyproject.toml), CI/CD workflows, README files, specs, and plans. These files could be manipulated to inject instructions that influence the skill's risk assessment or recommendation logic.
  • Ingestion points: Reads project manifests, documentation, and configuration files via cat, grep, and find.
  • Boundary markers: The instructions do not specify explicit delimiters or 'ignore' warnings for the agent when processing these files.
  • Capability inventory: The skill has the ability to write to the filesystem (.agents/skills, .claude/skills), execute shell commands, and perform network operations via external CLIs.
  • Sanitization: No specific sanitization or filtering of the content from ingested project files is mentioned before processing.
  • [DYNAMIC_EXECUTION]: The skill executes a bundled Python helper script (scripts/check_layout.py) to audit the skill directory structure and symlinks. It also notes that recommended skills may contain their own scripts for the agent to run.
  • [DATA_EXFILTRATION]: The skill reads and inventories the contents of the user's global skill directories (~/.agents/skills and ~/.claude/skills). While this is used for 'coverage' analysis, it exposes information about the user's global agent environment to the specific project context.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 5, 2026, 10:40 PM