The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The generateResponse action in convex/ai.ts creates an attack surface for indirect prompt injection by ingesting and processing external data.
Ingestion points: The prompt argument in the generateResponse handler (file: convex/ai.ts).
Boundary markers: None present; the prompt is interpolated directly into the LLM request body.
Capability inventory: The skill possesses network access capabilities via fetch and database write capabilities via ctx.runMutation (file: convex/ai.ts).
Sanitization: The implementation lacks input validation or output sanitization before storing the AI-generated response back into the database.
[EXTERNAL_DOWNLOADS]: The skill uses standard package management commands to install the convex library (npm install convex) and bootstrap projects (npm create convex@latest). These are standard operations for building with the Convex framework.
[DATA_EXFILTRATION]: The skill makes a network request to Anthropic's official API endpoint (https://api.anthropic.com/v1/messages). This is a well-known service and the request is used to provide the primary AI functionality described in the documentation.

convex-backend