The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to interact with external web pages, which serves as a vector for instructions embedded in data to influence agent behavior.
Ingestion points: The text, eval, and console commands in scripts/cdp.mjs extract data, exceptions, and console logs directly from the browser context.
Boundary markers: The skill includes a dedicated Safety section in SKILL.md that explicitly instructs the agent to treat page content as untrusted data and ignore instructions found within pages.
Capability inventory: The skill can navigate the browser, interact with elements (click, type), capture screenshots, and manage browser sessions via scripts/brave-cdp.sh.
Sanitization: The instructions require explicit human confirmation before submitting forms or sensitive actions and strictly forbid entering real credentials.
[COMMAND_EXECUTION]: The skill executes internal shell and Bun scripts to manage browser sessions and interact with the DevTools Protocol.
Evidence: The skill invokes scripts/brave-cdp.sh to start/stop the browser and scripts/cdp.mjs to send automation commands.
[DYNAMIC_EXECUTION]: The skill provides a command to execute arbitrary JavaScript within the target browser tab.
Evidence: The eval command in scripts/cdp.mjs uses the Runtime.evaluate protocol method to run code provided by the agent or user in the browser environment.

computer-user