The Agent Skills Directory

[PROMPT_INJECTION]: The skill mentions phrases such as "ignore previous instructions" and "disregard your task" within its security notice. These are not injection attempts, but defensive instructions designed to train the agent to recognize and ignore malicious directives found in untrusted external data during an audit. This represents a security best practice for agents processing third-party content.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves analyzing external website data, which presents an indirect prompt injection surface.
Ingestion points: The skill processes external content through the structure_artifacts, content_inventory, and user_research inputs, which may include fetched URL content and user-submitted text.
Boundary markers: The skill explicitly mandates the use of delimiter isolation (mentally scoping content within <untrusted-content> tags) and explicitly states that skill instructions take precedence over any directives found in the analyzed data.
Capability inventory: The skill is limited to analytical and reporting tasks. It does not utilize shell commands, file system writes, or network operations beyond standard URL retrieval for analysis.
Sanitization: The instructions guide the agent to disregard HTML/Markdown formatting and encoded characters intended to disguise instructions as content, effectively mitigating common injection techniques.

information-architecture-audit