The Agent Skills Directory

[PROMPT_INJECTION]: The skill mentions prompt injection bypass patterns exclusively within its defensive documentation. It provides clear instructions on how the agent should handle 'instruction-like' content found in web pages by treating it as data to be reported, not instructions to be executed.
[DATA_EXFILTRATION]: The skill provides robust guidelines to prevent data leakage, including the use of isolated browser contexts, avoiding the capture of session tokens, and managing credentials through environment variables rather than hardcoded strings.
[INDIRECT_PROMPT_INJECTION]: The skill models a comprehensive trust boundary, identifying web page content (DOM, console logs, network traffic) as untrusted. It explicitly forbids the agent from navigating to URLs or executing scripts found in this content without human confirmation.
[EXTERNAL_DOWNLOADS]: The skill references official Microsoft Playwright repositories and container images. These are trusted, well-known services for providing testing infrastructure.

web-testing