The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the mcporter command-line tool to manage the Chrome DevTools bridge. It executes subcommands to list browser tabs, restart the daemon, and call specific MCP tools.
[DATA_EXFILTRATION]: The instructions include a dedicated section for 'Secret Handling' which strictly prohibits the agent from outputting tokens, passwords, or other sensitive information harvested from the browser context.
[PROMPT_INJECTION]: The skill ingests untrusted data from external websites when taking snapshots or evaluating scripts, which constitutes an indirect prompt injection surface.
Ingestion points: Untrusted web content enters the context via take_snapshot and evaluate_script calls in SKILL.md.
Boundary markers: No specific delimiters or boundary warnings are provided to differentiate web content from system instructions.
Capability inventory: The skill allows for navigation, element interaction, and JavaScript execution, providing a significant range of actions that could be targeted by injection.
Sanitization: There is no evidence of sanitization or filtering of the data retrieved from pages before it is presented to the agent.
[REMOTE_CODE_EXECUTION]: The skill enables the execution of JavaScript within the browser context using the evaluate_script tool. This is the intended functional behavior for the described browser automation use case and originates from a reputable source.

browser-use