The Agent Skills Directory

[COMMAND_EXECUTION]: The wrapper script browser.js executes shell commands including playwright-cli and git. It forwards user/agent arguments directly to these tools.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to read and process content from a wide variety of external websites (Gmail, LinkedIn, Reddit, etc.). This content is untrusted and could contain instructions designed to influence the agent's behavior. The skill documentation includes warnings to 'paraphrase' rather than 'transcribe' to mitigate this, but the attack surface is significant.
[DATA_EXFILTRATION]: The skill provides explicit instructions and code patterns for extracting highly sensitive authentication data, including JWT tokens from localStorage, session cookies (like JSESSIONID), and CSRF tokens. It also provides a script in references/api-capture.md to intercept all network request bodies and headers. While intended for legitimate automation, these patterns expose credentials to the agent's context and output logs.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of the @playwright/cli package and performs periodic update checks by executing git fetch origin against its own source repository on GitHub. These operations target well-known and trusted services.
[DYNAMIC_EXECUTION]: The skill's core functionality relies on the exec eval command, which allows the agent to execute arbitrary JavaScript code within the context of a browser page. This is used for complex interactions, data extraction, and internal API calls.

browser-automation