The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides detailed guidance on exploiting browser-based agents by injecting instructions through untrusted data sources such as the DOM, accessibility trees, and screenshots. Mandatory Evidence Chain:\n- Ingestion points: The agent perceives web page content, rendered DOM nodes, and screenshots (SKILL.md).\n- Boundary markers: Documented as typically absent or insufficient for vision-based perception paths.\n- Capability inventory: The agent can navigate, fill forms, and operate across authenticated browser tabs (SKILL.md).\n- Sanitization: The analysis highlights that DOM-level sanitization fails to capture instructions smuggled via OCR or image pixels.\n- [DATA_EXFILTRATION]: Includes a self-contained Python probe script designed to serve as an out-of-band listener to verify if an agent successfully exfiltrates data (e.g., user emails or session tokens).\n- [PROMPT_INJECTION]: Contains examples of override instructions designed to hijack agent tasks, including commands embedded in HTML comments and hidden CSS elements.\n- [OBFUSCATION]: Documents the use of 1px text, hidden elements, and low-contrast text ('faint light-blue on yellow') to smuggle instructions past human review into OCR/vision channels.

hunt-agent-browser-use