external-llm-consulting

Fail

Audited by Gen Agent Trust Hub on Mar 24, 2026

Risk Level: HIGHCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute external CLI commands (codex and gemini). User-provided prompts and codebase context are interpolated directly into shell command strings (e.g., codex exec ... "<prompt>"), which is vulnerable to command injection if the input contains shell metacharacters.
  • [EXTERNAL_DOWNLOADS]: The skill provides instructions to download and install external Node.js packages globally (npm install -g) if the tools are not found on the system. It also references a remote repository for the Gemini CLI (github.com/google-gemini/gemini-cli) to provide documentation.
  • [REMOTE_CODE_EXECUTION]: The recommended installation commands target unverified and suspicious packages. Specifically, @anthropic-ai/gemini-cli is highly suspect as Gemini is a Google product, not an Anthropic one; this naming convention is consistent with malicious package impersonation or supply chain attacks. Similarly, @openai/codex is a non-standard package name for OpenAI's official tooling.
  • [DATA_EXFILTRATION]: The skill's primary workflow involves gathering sensitive codebase context from files such as CLAUDE.md and AGENTS.md, which is then transmitted to external third-party LLM services. While the skill advises against including secrets, the automated collection and transmission of architectural patterns and constraints to external endpoints presents a data exposure risk.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Mar 24, 2026, 09:57 PM