The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs users to install software using curl -fsSL https://agent-browser.sh | bash. This method executes a remote script directly in the shell without any verification, allowing for arbitrary code execution on the user's machine.
[EXTERNAL_DOWNLOADS]: The skill references multiple external sources for downloading executable content, including the domain agent-browser.sh and the NPM registry. Additionally, the command agent-browser skills get core dynamically fetches workflow instructions from a remote source at runtime.
[COMMAND_EXECUTION]: The skill provides instructions for the agent to execute complex CLI commands, including navigate, click, fill, and screenshot. These commands interact with browser processes and have the capability to read and modify local configuration files like ~/.agent-browser/config.toml.
[PROMPT_INJECTION]: There is a significant discrepancy in the skill's attribution. While the metadata and description claim the author is 'Vercel Labs', the skill is provided by 'peter-tu-zynkr' and references 'zynkr.ai'. This deceptive use of a well-known organization's name to distribute software is a major security red flag.
[DATA_EXFILTRATION]: The CLI tool is designed to extract data from websites, including accessibility trees, HTML, and screenshots. While this is the intended functionality, it also creates a surface for indirect prompt injection and the potential for sensitive data to be transmitted to external endpoints via the 'chat' gateway.

agent-browser