agent-token-efficiency

Warn

Audited by Gen Agent Trust Hub on Jul 24, 2026

Risk Level: MEDIUMREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The installation guide in references/tools-install.md includes a command for the 'Caveman' tool that downloads a PowerShell script from a remote URL (https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.ps1) and executes it immediately using the iex (Invoke-Expression) command, bypassing local security review.
  • [EXTERNAL_DOWNLOADS]: The skill provides instructions to download and install several third-party components from unverified GitHub repositories and package registries, specifically DietrichGebert/ponytail, JuliusBrussee/caveman, and the headroom-ai Python package.
  • [COMMAND_EXECUTION]: The skill promotes the use of 'Headroom' (headroom wrap cursor), which acts as a local proxy to intercept and modify traffic between the AI agent and the model provider. This introduces a significant trust dependency on unverified third-party software that handles all model communication.
  • [COMMAND_EXECUTION]: The skill utilizes npx skills add commands to fetch and install external modules from various GitHub accounts, which can lead to the execution of unverified remote code in the local environment.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 24, 2026, 02:47 AM
Security Audit — agent-trust-hub — agent-token-efficiency