The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes patterns for creating agents that ingest entity data from the platform. This data is often used directly in prompts, creating a vulnerability surface for indirect prompt injection attacks where malicious data in an entity (like a task description or return request) could influence the agent's behavior.
Ingestion Point: Documentation in references/web-spa.md and references/web-tanstack.md shows agents reading from entities such as ReturnRequest or Order via platform SDKs.
Boundary Markers: The provided code examples interpolate data directly (e.g., via JSON.stringify(request)) without using delimiters or system-level instructions to ignore embedded commands.
Capability Inventory: Agents are configured with operations like 'update' and 'delete' on entities, and the ability to trigger backend functions or server routes, which can be misused if an injection occurs.
Sanitization: No sanitization, escaping, or input filtering is demonstrated for entity-sourced data before it enters the model context.
[EXTERNAL_DOWNLOADS]: The skill documentation references several external packages required for building and running agent loops. These resources are from the vendor or recognized well-known libraries.
Dependencies: Fetches @base44/sdk (vendor resource), ai and @ai-sdk/openai-compatible (well-known Vercel libraries), and standard libraries like zod and @tanstack/ai.

base44-platform-agents