The Agent Skills Directory

[PROMPT_INJECTION]: The file references/trust-boundaries.md contains text patterns commonly used in prompt injection attacks (e.g., "ignore previous instructions") as part of its defensive documentation. These are provided as examples to teach the agent to identify and disregard instructions embedded in untrusted web content.
[REMOTE_CODE_EXECUTION]: The skill documents an installation method for a companion utility in references/authentication.md using curl -fsSL https://raw.githubusercontent.com/leeguooooo/bitwarden-use/main/install.sh | sh. This script is hosted on the vendor's own repository and is used for credential management integration.
[COMMAND_EXECUTION]: The skill leverages the chrome-use CLI via the Bash tool to perform browser interactions. This includes the eval command, which executes JavaScript in the browser context, and various file-system operations for state management (state save/state load).
[DATA_EXFILTRATION]: While the skill accesses sensitive browser data such as cookies and session storage, it includes a dedicated trust-boundaries.md reference that provides strict guidelines to the agent against echoing or exfiltrating these secrets.

core