The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill audit process ingests untrusted data from web pages, providing a surface for indirect prompt injection.\n
Ingestion points: Scripts in the scripts/ directory, such as Interactions.js and Layout-Shift-Loading-and-Interaction.js, read data from browser performance APIs and the DOM.\n
Boundary markers: There are no boundary markers used to separate audited content from agent instructions in the returned data.\n
Capability inventory: The agent can execute scripts and read console output from the browser context, which includes access to all page data.\n
Sanitization: Element selectors, script URLs, and other page attributes are not sanitized before being returned to the agent.\n- [COMMAND_EXECUTION]: The skill provides scripts that deeply instrument the browser environment by monkey-patching standard prototypes.\n
Evidence: scripts/Forced-Synchronous-Layout.js patches setAttribute and various layout properties of Element.prototype. scripts/Scroll-Performance.js patches EventTarget.prototype.addEventListener.

webperf-interaction