The Agent Skills Directory

[PROMPT_INJECTION]: The skill documents an interface for processing external data through a language model which is susceptible to indirect prompt injection. * Ingestion points: The session.respond(to:) and session.streamResponse(to:) methods in SKILL.md accept strings that may contain untrusted data. * Boundary markers: The examples do not demonstrate the use of delimiters or system instructions to distinguish between developer commands and user-provided data. * Capability inventory: The skill introduces a Tool protocol in SKILL.md which allows the model to trigger external actions, such as network requests as seen in the WeatherTool example. * Sanitization: There is no documentation or guidance regarding the sanitization, validation, or escaping of input data before it is passed to the model session.

on-device-ai