The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the development of applications and MCP tools that ingest untrusted data from external sources, including HTTP requests, file uploads, and model-provided tool arguments. These inputs are processed by applications that have access to significant capabilities such as database operations via AbstractData, outbound networking via URLRequest, and file system access for uploads. While the skill includes explicit warnings and examples of input validation to mitigate these risks, the inherent architecture creates a surface for indirect prompt injection attacks.
Ingestion Points: Found in SKILL.md (MCP tool arguments) and references/system-usage.md (HTTP request parameters, session data, and file uploads).
Boundary Markers: The skill suggests using validation logic (length, regex) in SKILL.md but does not define specific boundary delimiters for prompt interpolation.
Capability Inventory: Documented across all files, including database CRUD operations (references/database.md), file system writes (references/system-usage.md), outbound HTTP requests (references/system-usage.md), and CLI command execution via the dispatcher (references/architecture.md).
Sanitization: The skill provides a clear example of input sanitization using regex and length checks in the MCP tool section of SKILL.md.

tinystruct-patterns