The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides run-code and eval commands that allow the agent to execute arbitrary JavaScript logic within the browser context and the local execution environment, facilitating the dynamic execution of unverified code strings.
[CREDENTIALS_UNSAFE]: The skill features tools to extract and manage sensitive browser session data, including cookies, localStorage, and session state, which can be saved to local files like auth.json, potentially exposing authentication tokens.
[COMMAND_EXECUTION]: The documentation includes a recommended builder.config.json that grants the agent access to a wide range of powerful shell commands with wildcards, such as curl, rm, and chmod, which significantly broadens the potential impact of any command execution.
[EXTERNAL_DOWNLOADS]: The skill relies on external software from the NPM registry, including the installation of @playwright/cli and execution via npx, which are fetched from well-known services.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it processes content from external, untrusted websites through browser navigation. Ingestion points: Browser navigation via open and goto commands. Boundary markers: Absent. Capability inventory: File writes (state-save), script execution (run-code, eval), and interactive element control. Sanitization: No sanitization or filtering of DOM content is mentioned.

playwright-cli