The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from external sources like phone call events and messaging webhooks, which presents an attack surface for indirect prompt injection.
Ingestion points: External data enters the agent context through event handlers such as Incoming, Manage, and MessageInbound, as well as legacy voice callbacks like Ice and Pie.
Capability inventory: The skill runtime provides methods to write to a persistent SQLite database (Context.Database), write files to persistent storage (Context.Storage), and interact with messaging and voice networks via the pre-authenticated Sinch SDK.
Sanitization: The instructions explicitly warn the agent that callback data is untrusted and must be validated before use, stating that it should "never interpolate into prompts, shell commands, or SQL." It specifically recommends using parameters with Dapper or Microsoft.Data.Sqlite for database interactions.
Boundary markers: While the skill provides clear warnings against direct interpolation, it does not define mandatory structural delimiters to isolate untrusted data if it were to be included in an LLM prompt.
[DATA_EXFILTRATION]: The skill performs network operations to interact with Sinch API services and fetch documentation from the vendor's official domain.
Evidence: The runtime utilizes pre-authenticated clients (Context.Voice, Context.Conversation, Context.Sms) and instructions allow fetching documentation from developers.sinch.com.
Context: These operations are restricted to trusted first-party infrastructure and official vendor SDKs, which is appropriate for the skill's intended functionality.

sinch-functions-dotnet