us-program-research
Pass
Audited by Gen Agent Trust Hub on Sep 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from various external web sources, which creates a surface for indirect prompt injection attacks.
- Ingestion points: The skill retrieves student reviews and program data from websites including Niche, Reddit, GMAT Club, GradCafe, and various university domains, as specified in
references/research-sources.mdand used in the Phase 3 subagent prompts inreferences/subagent-prompts.md. - Boundary markers: The subagent prompts in
references/subagent-prompts.mduse structured extraction fields but lack explicit safety delimiters or instructions to ignore potential commands embedded within the fetched web content. - Capability inventory: The skill has the capability to write a generated markdown report to the local filesystem (
SKILL.mdPhase 5) and spawn sub-tasks using the platform's Task tool. - Sanitization: There is no mention of sanitization or filtering of the ingested external content before it is processed and included in the final report.
Audit Metadata