gco-research

Warn

Audited by Gen Agent Trust Hub on Jun 25, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: User-provided content, such as the research topic and slug, is interpolated directly into shell commands. In Step 3, the instruction executes bash $HOME/.claude/skills/gco/scripts/gco-run.sh "<research prompt>". If the user input contains shell metacharacters and the agent fails to escape them, this allows for arbitrary command execution.
  • [EXTERNAL_DOWNLOADS]: In Step 7, the skill executes pnpm dlx @takazudo/mdx-formatter. This command fetches and runs code from the public NPM registry at runtime. While the package belongs to the author, executing unpinned code from a public registry introduces a supply chain risk.
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It ingests data from the web via Copilot output or a researcher subagent and synthesizes the findings. The instructions lack boundary markers or explicit directives to ignore instructions embedded within the researched content.
  • Ingestion points: Copilot output logs and subagent research results in SKILL.md.
  • Boundary markers: None identified.
  • Capability inventory: Shell access (bash), Node.js execution, and file system writes.
  • Sanitization: None identified.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jun 25, 2026, 07:18 PM
Security Audit — agent-trust-hub — gco-research