The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the processing of incoming messages from external communication channels (SMS, WhatsApp, Voice) which are subsequently passed to an LLM. This architecture creates an inherent attack surface for indirect prompt injection.
Ingestion points: User messages are ingested via the user_message (Python) and message (TypeScript) parameters within the respective on_message_ready callbacks in the provided code snippets.
Boundary markers: The quickstart examples demonstrate direct interpolation of untrusted user messages into the conversation history or messages array without the use of explicit delimiters or instructions to ignore embedded commands.
Capability inventory: The TAC framework provides tools for memory retrieval, semantic search in knowledge bases, and Studio Flow handoff (human escalation). It also supports custom tool creation, such as the send_email example provided in the documentation.
Sanitization: While the documentation mentions that channels validate events, the core logic shown in the examples does not implement specific sanitization or filtering of the user-provided text before it reaches the LLM.
[EXTERNAL_DOWNLOADS]: The skill fetches SDKs and integration packages from Twilio's official GitHub repositories (e.g., github.com/twilio/twilio-agent-connect-python).
[CREDENTIALS_UNSAFE]: The skill requires the configuration of sensitive credentials, including TWILIO_ACCOUNT_SID, TWILIO_AUTH_TOKEN, TWILIO_API_KEY, and TWILIO_API_SECRET. The documentation follows security best practices by recommending the use of environment variables and advising against committing these secrets to version control.

twilio-agent-connect