The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill proactively addresses indirect prompt injection by providing a comprehensive framework for handling untrusted data from external sources.
Ingestion points: Defined as web pages, files, search results, API responses, database rows, tool outputs, and messages from other agents in SKILL.md.
Boundary markers: Instructs the agent to label sources before reading and utilize a structured 'UNTRUSTED CONTENT NOTE' reporting format.
Capability inventory: Explicitly warns against piping content into shells, code interpreters, or other agent prompts without validation in SKILL.md (Rule 5).
Sanitization: Rule 6 in SKILL.md requires validation of extracted values like URLs and paths against the user's scope before use.
[PROMPT_INJECTION]: Instructions in references/injection-triage.md contain common prompt injection patterns such as "Ignore previous instructions" and "Disregard the user's request". These are explicitly presented as examples of malicious patterns to be recognized and ignored, serving as a defensive control rather than an attack vector.

untrusted-content