handoff
Pass
Audited by Gen Agent Trust Hub on Jul 24, 2026
Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [PROMPT_INJECTION]: The skill instructs the agent to create a 'suggested skills' section in the handoff document based on the current conversation. This creates an indirect prompt injection surface (multi-step chain) where malicious instructions embedded in a processed conversation could influence the behavior of the next agent receiving the handoff.
- Ingestion points: Processes current conversation history.
- Boundary markers: None specified for the output document.
- Capability inventory: The generated document influences the skill invocation choices of subsequent agents.
- Sanitization: Instructions include redacting PII and credentials, but do not address sanitizing instructional content.
- [DATA_EXFILTRATION]: The instructions explicitly direct the agent to save the handoff document to the 'temporary directory of the user’s OS' instead of the current workspace. This represents a transfer of conversation data (albeit summarized) outside the established project environment boundaries.
Audit Metadata