The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection as it facilitates the ingestion of external web content (via 'agent-browser open', 'get text', and 'snapshot'). * Ingestion points: 'agent-browser' commands that read page titles, body text, or accessibility snapshots. * Boundary markers: The skill advises using targeted extraction ('get text @e#') instead of full-page dumps to limit the amount of untrusted content processed at once. * Capability inventory: The skill allows for complex browser interactions including form filling, element clicks, and script execution ('eval'). * Sanitization: The instructions emphasize 'trusting page content without validation' as an anti-pattern and recommend verification before processing results.- [DYNAMIC_EXECUTION]: The tool supports 'agent-browser eval' and 'agent-browser batch', which allow for the execution of arbitrary JavaScript in the browser or batching of shell commands. These are documented as powerful features with corresponding security warnings in the rules.- [COMMAND_EXECUTION]: The skill utilizes local CLI tools ('agent-browser', 'portless') and provides commands for host clipboard interaction ('clipboard read/write'), which are identified as potential privacy risks in sensitive environments.- [EXTERNAL_DOWNLOADS]: The documentation references the installation of 'portless' and '@vercel/sandbox' from official registries. These are well-known technology services and are considered safe under the trust policy.

browser-tools