The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill inherently processes untrusted data from target applications via the browser's DOM, console logs, and network responses during its various probes.
Ingestion points: Data enters the agent's context through page.evaluate calls, console listeners (console-network.md), and network interception (failure-injection.md, layout-shift.md).
Boundary markers: The instructions mandate the capture and storage of raw evidence artifacts (screenshots, JSON logs) to enable human review of the agent's findings, providing a verification layer against potentially misleading data.
Capability inventory: The skill has the capability to perform network requests and write artifacts to the local file system (artifactDir).
Sanitization: The skill focuses on extracting structural, performance, and accessibility metrics rather than interpreting or re-executing text content, which minimizes the exploitability of the injection surface.
[DYNAMIC_EXECUTION]: The skill utilizes page.evaluate() and page.addInitScript() to execute specialized JavaScript recipes within the browser context. These scripts are used for tasks like measuring hit targets, tracing focus, and observing layout shifts. This is a standard and necessary implementation for browser-based UI verification and is controlled by the skill's predefined recipes.

ui-verification