agent-handoff-protocols
Pass
Audited by Gen Agent Trust Hub on Aug 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill documents architectural patterns for agent communication that introduce a vulnerability surface for indirect prompt injection. Specifically, the
summarizedTransferstrategy interpolates unsanitized context data into a prompt for a secondary agent.\n- Ingestion points:HandoffPackageandContextobjects described inSKILL.md.\n- Boundary markers: The provided pseudo-code lacks delimiters (e.g., XML tags or triple quotes) to isolate untrusted context from instructions.\n- Capability inventory: The patterns involve passing data to other agents (summarizeAgent.run) and message queues (this.queue.enqueue), extending the potential impact of an injection.\n- Sanitization: No validation, escaping, or filtering of the transferred context is included in the design templates.
Audit Metadata