The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection as it is designed to navigate to and analyze untrusted external web content.
Ingestion points: The skill uses browser_navigate, browser_snapshot, and browser_vision to ingest data from user-provided target URLs in SKILL.md.
Boundary markers: There are no instructions or delimiters in the skill files that direct the agent to ignore or isolate commands found within the web pages being tested.
Capability inventory: The skill allows the agent to interact with the target pages via browser_click, browser_type, and browser_press, and it writes reports and screenshots to the local filesystem using templates provided in templates/dogfood-report-template.md.
Sanitization: No evidence of content validation or sanitization is present before the agent processes the retrieved data.

dogfood