care-gap-closure
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted information from patient records and provider notes, creating a surface for indirect instructions to influence the agent.
- Ingestion points: Untrusted data is ingested through
patient_dataobjects in Python logic, patient response fields in the communication scripts, and theNotessection in theCARE_GAP_CLOSUREandProvider Escalationtemplates. - Boundary markers: There are no explicit boundary markers or delimiters defined in the templates (such as
CARE_GAP_CLOSUREor the escalation email) to isolate potentially untrusted content from the agent's instructions. - Capability inventory: The skill is designed to facilitate interactions with patient portals, phone systems, and scheduling workflows, which could be misused if an injection is successful.
- Sanitization: The skill does not implement or recommend sanitization or filtering for external data before it is included in reports or used to generate communication.
Audit Metadata