search-first

Pass

Audited by Gen Agent Trust Hub on May 14, 2026

Risk Level: SAFEDATA_EXFILTRATIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill instructs the agent to read platform configuration files (~/.claude/settings.json) and the local skills directory (~/.claude/skills/). While intended to discover existing tools and prevent duplication, these files can contain sensitive environment metadata and configuration settings.
  • [EXTERNAL_DOWNLOADS]: The workflow explicitly promotes searching for and installing third-party packages from public registries (npm, PyPI) and GitHub. This behavior facilitates the introduction of unvetted external code into the local development environment, posing a supply chain risk.
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its reliance on untrusted external data sources during the research phase.
  • Ingestion points: The agent processes search results, repository descriptions, and documentation from npm, PyPI, GitHub, and general web queries.
  • Boundary markers: No specific delimiters or "ignore instructions" directives are provided in the subagent task prompt template to isolate untrusted data from the agent's instructions.
  • Capability inventory: The workflow leads directly to package installations (npm install, pip install) and the configuration of MCP servers based on information retrieved from the web.
  • Sanitization: There is no mechanism described for validating or sanitizing external content before it is used to inform the agent's decisions in the "Decision Matrix."
Audit Metadata
Risk Level
SAFE
Analyzed
May 14, 2026, 10:51 AM