The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external landing pages via scripts/fetch_page.py and scripts/analyze_landing.py, which is an inherent part of its audit functionality.
Ingestion points: The agent ingests data extracted from the Document Object Model (DOM) of user-provided URLs, including H1 tags, page titles, and general text content.
Boundary markers: The instructions do not explicitly define delimiters to separate extracted web content from agent instructions, though the data is handled within specific audit workflows.
Capability inventory: The skill orchestrates subagents using the spawn_agent capability and performs network read operations. It includes limited local file-writing for saving screenshots in scripts/capture_screenshot.py but lacks arbitrary code execution or credential access.
Sanitization: Extracted web content is processed in its raw form for analysis without specific sanitization for LLM-based injection.
[COMMAND_EXECUTION]: The skill uses Python scripts to perform deep landing page analysis and visual audits. These scripts implement robust safety logic (_is_public_host and _is_public_ip) to validate that all requested URLs resolve to public internet addresses, effectively mitigating potential SSRF (Server-Side Request Forgery) attacks.
[EXTERNAL_DOWNLOADS]: The skill's scripts utilize well-known and trusted Python packages, including playwright for browser automation and requests for HTTP communication, to facilitate its auditing tasks.

ads