agy-harness

Pass

Audited by Gen Agent Trust Hub on Oct 11, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPERSISTENCEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill provides instructions to download and install the Antigravity CLI tool directly from Google's official package repository using the Windows Package Manager (winget).
  • [COMMAND_EXECUTION]: The skill instructions involve executing the winget installer, the agy.exe binary, and various system utilities such as where.exe to verify the installation and drive CLI tasks.
  • [PERSISTENCE]: The documentation describes the agy install command, which is used to append the tool's binary path and shell aliases to the user's shell profile, ensuring the environment remains configured across sessions.
  • [INDIRECT_PROMPT_INJECTION]: The skill creates a multi-step execution chain by allowing a host agent to pass arbitrary natural language prompts to the Gemini CLI (agy --print). This creates a vulnerability surface if the sub-agent is configured with the recommended high-privilege settings.
  • Ingestion points: The prompt argument passed to agy --print in SKILL.md serves as the entry point for untrusted data into the sub-agent's context.
  • Boundary markers: There are no explicit delimiters or instruction-guarding markers provided to separate the system instructions from the content of the prompt passed to the CLI.
  • Capability inventory: The skill explicitly recommends configuring the tool's settings (settings.json) with read_file(*) and command(*) permissions, which allows the sub-agent to read local files and execute arbitrary shell commands.
  • Sanitization: The harness does not implement any input sanitization or validation of the sub-agent's output before it is processed by the host environment.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 11, 2026, 11:55 AM