The Agent Skills Directory

[SAFE]: The skill implements a strong security posture in SKILL.md called the 'Untrusted content boundary'. This instructs the agent to treat all retrieved material (text, HTML, metadata, documents) as untrusted data rather than instructions, helping to prevent the execution of malicious commands embedded in external content.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external content, which is a potential surface for indirect prompt injection.
Ingestion points: Third-party text, HTML, API responses, and documents retrieved during the verification process (specified in SKILL.md).
Boundary markers: The skill explicitly mandates the use of <EXTERNAL_DATA source="..."> delimiters to isolate untrusted content from the agent's core instructions.
Capability inventory: The skill provides a Python snippet in references/verification-trail-and-archiving.md that uses the requests library for network operations. It does not contain capabilities for arbitrary command execution or writing to sensitive system files.
Sanitization: The skill instructions include explicit requirements to ignore embedded requests to run tools or reveal secrets, validate all URLs/redirects, and cap parsing depth and content size.
[EXTERNAL_DOWNLOADS]: A reference implementation in references/verification-trail-and-archiving.md uses the Python requests library to automate web archiving to well-known and trusted services, including the Internet Archive (Wayback Machine), Archive.today, and Perma.cc. These network operations are limited to these specific archival targets and are used for legitimate evidence preservation.

source-verification