The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the processing of external and untrusted data, which presents an indirect prompt injection vulnerability surface.
Ingestion points: Untrusted data enters the agent context through the prompt() method in the SalesCoach example, as well as via Transcription::fromStorage() and Document::fromPath() operations.
Boundary markers: The provided implementation examples lack explicit boundary delimiters or instructions to ignore embedded commands within the processed data.
Capability inventory: The documented SDK features support network operations (via AI providers), file system interactions through the Files and Stores modules, and the execution of arbitrary logic via the HasTools interface.
Sanitization: The skill's examples do not demonstrate input sanitization, schema validation for incoming data, or escaping mechanisms before interpolating external content into prompts.

ai-sdk-development