ai-core/custom-backend-integration
Pass
Audited by Gen Agent Trust Hub on Sep 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill documents methods for ingesting and displaying content from external backend connections, which establishes a potential surface for indirect prompt injection.
- Ingestion points: The
useChathook and various connection adapters (SSE, NDJSON, and WebSockets) defined inSKILL.mdingest untrusted data from remote API endpoints. - Boundary markers: The provided code examples and documentation do not incorporate delimiters or specific instructions to ignore embedded commands within the external data stream.
- Capability inventory: The documented patterns in
SKILL.mdutilizefetchandWebSocketfor network communication and include asendMessagecapability for interaction. - Sanitization: The UI rendering examples directly render message parts (
part.content) into the DOM without demonstrating sanitization or filtering of potential control tokens or malicious instructions.
Audit Metadata