The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides the browser_evaluate tool, which allows the agent to execute arbitrary JavaScript code within the browser page context. It also includes browser_add_script_tag for injecting external scripts and browser_wait_for_function for running logic during polling.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process content from untrusted external websites via tools like browser_goto and browser_get_text, creating a significant surface for indirect prompt injection attacks. \n
Ingestion points: Any website navigated to or read by the agent (e.g., via browser_goto, browser_get_text, browser_get_html). \n
Boundary markers: No instructions or delimiters are provided to help the agent distinguish between its core instructions and instructions embedded in untrusted web data. \n
Capability inventory: The skill offers extensive browser control including JS execution (browser_evaluate), network manipulation (browser_mock_response), cookie/storage access, and file interactions. \n
Sanitization: No explicit sanitization, filtering, or validation of external web content is described.
[DATA_EXFILTRATION]: Several tools facilitate the extraction of sensitive information. browser_get_cookies and browser_get_local_storage retrieve session data, while browser_storage_state allows saving the entire browser state, including authentication cookies, to a local file path.
[COMMAND_EXECUTION]: The skill provides comprehensive control over the browser environment and local file interactions (e.g., browser_upload_file, browser_screenshot, browser_pdf), which can be leveraged for malicious purposes if the agent is manipulated by untrusted input.

playwright-browser