The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The script scripts/extract-cookies.py is designed to harvest authentication cookies from local browser profiles, targeting paths such as ~/Library/Application Support/Google/Chrome/Default/Cookies and similar directories for Brave and Zen browsers.
[CREDENTIALS_UNSAFE]: The skill attempts to retrieve encryption keys for browser cookie storage by executing the security find-generic-password command to access the macOS Keychain.
[COMMAND_EXECUTION]: Multiple components use subprocess.run to execute shell commands, including lsof and kill in scripts/cloak-launch.py to manage browser processes, and node in engine/executor.py to run JavaScript templates.
[REMOTE_CODE_EXECUTION]: Setup instructions in references/insane-search/playwright.md and references/agent-browser-setup.md direct the user to execute commands that download and run code from remote registries, such as npx @playwright/mcp@latest.
[EXTERNAL_DOWNLOADS]: The skill's components and documentation encourage the installation of several third-party dependencies and CLI tools from external sources, including yt-dlp, curl_cffi, and specialized stealth Chromium binaries.
[PROMPT_INJECTION]: The skill ingests untrusted data from the public web via its fetching engine in engine/fetch_chain.py. It implements a defensive evidence chain in engine/content_safety.py using hash-based delimiters ([BEGIN UNTRUSTED WEB CONTENT]) and a signal scanner to detect instruction overrides or credential access patterns in the fetched content. Boundary markers and sanitization reports are used to mitigate the risk of the agent obeying instructions embedded in web pages.
[DATA_EXFILTRATION]: The skill establishes a potential exfiltration path by combining the capability to read sensitive local files (browser cookies and Keychain data) with outbound network operations performed via the curl_cffi transport layer.

browsing