record-reconciliation
Pass
Audited by Gen Agent Trust Hub on Oct 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from sources like emails, transcripts, and chat logs, which creates a potential surface for indirect prompt injection. However, it implements a comprehensive 'Untrusted content boundary' contract to mitigate these risks:
- Ingestion points: Data is ingested from emails, transcripts, chat, briefings, or third-party records (specified in
SKILL.md). - Boundary markers: The skill mandates the use of
<EXTERNAL_DATA>tags to clearly delimit untrusted content. - Capability inventory: The skill performs database searches, record writes, and URL fetching.
- Sanitization: It explicitly instructs the agent to ignore embedded instructions (e.g., requests to run tools or reveal secrets) and requires URL validation to reject private network destinations.
Audit Metadata