The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides a framework for processing external data which creates a vulnerability surface for indirect prompt injection.
Ingestion points: The documentation in references/getting-started.md demonstrates that the agent ingests untrusted user input via the LanguageModelSession.respond(to:) method.
Boundary markers: While the skill advises defining the model's role in instructions, it does not explicitly define or enforce boundary markers or delimiters for the user-supplied prompt content.
Capability inventory: The references/tool-calling.md file demonstrates tools with significant capabilities, including network access via WeatherService and access to sensitive PII through a FindContactTool that reads contact data.
Sanitization: The skill includes a proactive safety warning in references/getting-started.md advising developers never to interpolate user input into system instructions to prevent prompt injection, which partially mitigates the risk.

foundation-models