The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill connects to the user's primary Chrome browser instance via the Chrome DevTools Protocol (CDP). This allows the agent to inherit and use the user's active login sessions, cookies, and saved credentials on any website it visits.
[COMMAND_EXECUTION]: The skill implements a local HTTP proxy (scripts/cdp-proxy.mjs) that allows for the execution of arbitrary JavaScript code within the browser context via the /eval endpoint.
[DATA_EXFILTRATION]: The /setFiles API endpoint in the CDP proxy allows the agent to programmatically select local file paths to be uploaded to web forms. This capability could be exploited by indirect prompt injections on malicious websites to trick the agent into exfiltrating sensitive local files such as SSH keys, configuration files, or environment variables.
[COMMAND_EXECUTION]: The SKILL.md file utilizes dynamic context injection (the !command`` syntax) to execute shell commands at runtime when the skill is loaded, specifically using node -e to list directory contents.
[EXTERNAL_DOWNLOADS]: The skill integrates with r.jina.ai, an external third-party service, to convert web pages into Markdown for processing.

web-access