The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill interacts with external, untrusted websites, creating an inherent surface for indirect prompt injection attacks.\n- Ingestion points: The run_web_automation tool visits and processes content from arbitrary URLs provided at runtime.\n- Boundary markers: Instructions in SKILL.md define a behavioral trust boundary by explicitly warning the agent that page content is untrusted and should not be followed as instructions.\n- Capability inventory: The skill provides tools for active browser interaction, including clicking, typing, and submitting forms, which are high-capability actions.\n- Sanitization: The skill mitigates risks by requiring user confirmation for sensitive actions and instructing the agent to treat page content strictly as data, filtering out potential injection attempts.

tinyfish-automation