gemini-researcher

Pass

Audited by Gen Agent Trust Hub on Feb 26, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes the pbcopy command via the shell to copy generated text to the macOS clipboard. It correctly utilizes a quoted heredoc (<<'PROMPT') to prevent the shell from interpreting any special characters within the generated content, which is a security best practice for handling dynamic text in shell commands.
  • [PROMPT_INJECTION]: The skill demonstrates an indirect prompt injection surface because it interpolates untrusted user data into a prompt intended for another LLM (Gemini Deep Research).
  • Ingestion points: User-provided research requests enter the agent context in the instructions defined in SKILL.md.
  • Boundary markers: No explicit delimiters or instructions to ignore embedded commands are included in the generated prompt.
  • Capability inventory: The skill is capable of executing the pbcopy command via shell.
  • Sanitization: No sanitization or escaping of the user-provided research request is performed before interpolation. While this creates a path for indirect injection into the secondary LLM session, the skill itself possesses no sensitive permissions or network capabilities that could be compromised.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 26, 2026, 05:18 PM