The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is centered around a CLI tool (browser-use) that performs a wide range of browser actions, including session management and interactive element control.\n- [DYNAMIC_EXECUTION]: The skill provides tools for executing arbitrary code at runtime to handle complex automation logic.\n
Evidence: The browser-use eval command allows execution of JavaScript code within the browser context to interact with page content or extract data (found in SKILL.md).\n
Evidence: The browser-use python command allows for the execution of Python statements to directly interface with the Chrome DevTools Protocol (CDP) for low-level browser control (found in references/cdp-python.md).\n- [INDIRECT_PROMPT_INJECTION]: As a tool designed to read and interact with web content, the skill is susceptible to indirect prompt injection where a malicious website could attempt to influence the agent's behavior.\n
Ingestion points: External web content enters the agent's context through commands like browser-use state, browser-use get text, and browser-use get html (found in SKILL.md).\n
Boundary markers: The skill incorporates a mandatory safety reference, references/untrusted-content-rule.md, which explicitly instructs the agent to treat fetched text as data and never follow instructions found within it.\n
Capability inventory: The agent possesses capabilities to execute code (eval, python), upload files (upload), manage cookies (cookies set/export), and create network tunnels (tunnel), which could be targets for an injection attack.\n
Sanitization: The skill relies on model-level guidance (the untrusted-content rule) rather than programmatic sanitization of web content.\n- [EXTERNAL_DOWNLOADS]: The skill interacts with external services and downloads updates from vendor infrastructure.\n
Evidence: The browser-use tunnel command utilizes Cloudflare's tunneling service to expose local ports, which is a well-known and trusted technology service.\n
Evidence: The browser-use profile update command downloads binary updates, presumably from the author's official repository, which is a standard vendor resource.\n- [DATA_EXFILTRATION]: Commands for exporting sensitive browser data and uploading local files provide a potential path for data exfiltration if misused.\n
Evidence: The browser-use cookies export command allows saving browser cookies to a local file, exposing sensitive session data.\n
Evidence: The browser-use upload command enables the agent to pick local files and upload them to web forms on arbitrary websites.

browser-use