The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill contains automated logic to extract and decrypt sensitive browser session cookies directly from the host machine. The implementation in src/cookie-import-browser.ts targets local cookie databases for browsers including Chrome, Arc, Brave, and Edge, and utilizes the macOS Keychain to retrieve storage encryption keys.
[COMMAND_EXECUTION]: The skill executes external system utilities using Bun.spawn. Specifically, it calls the security command to interact with the macOS Keychain (src/cookie-import-browser.ts) and the open command to launch the cookie picker UI in the default browser (src/write-commands.ts).
[DATA_EXFILTRATION]: The skill facilitates the exposure of highly sensitive authentication data. The cookie-import-browser command allows the agent to read and import browser cookies into its current session, which could lead to session hijacking of the user's accounts if the agent is manipulated by a malicious website.
[REMOTE_CODE_EXECUTION]: The skill allows the execution of arbitrary JavaScript from files on the host system. The eval command in src/read-commands.ts reads content from local file paths and executes it within the browser context via page.evaluate(). While restricted to specific directories, this could be chained with other agent capabilities to execute locally-generated malicious code.
[PROMPT_INJECTION]: The skill possesses a significant indirect prompt injection surface as it is designed to ingest and process untrusted content from the web. Information extracted via text, html, or accessibility commands (src/read-commands.ts) is returned to the agent without sanitization, potentially allowing malicious webpages to influence the agent's behavior and trigger high-privilege commands.

browse