us-program-research

Pass

Audited by Gen Agent Trust Hub on May 10, 2026

Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection by ingesting untrusted student reviews from external websites.
  • Ingestion points: The skill performs searches on public platforms such as Reddit, Niche, and GMAT Club to gather student feedback (documented in references/subagent-prompts.md).
  • Boundary markers: There are no explicit instructions or delimiters used to separate the fetched content from the agent's internal instructions.
  • Capability inventory: The skill utilizes web search tools, subagents, and file-writing capabilities to produce a Markdown report (SKILL.md).
  • Sanitization: The instructions do not define methods for sanitizing or validating the content retrieved from external sources.
  • [DATA_EXFILTRATION]: The skill initiates network operations to perform research and collect reviews from well-known external web services (Reddit, Niche, and university domains).
  • [DATA_EXFILTRATION]: The skill collects user-provided personally identifiable information (PII), such as name, email address, and academic performance metrics, to create a customized application action plan for the user.
Audit Metadata
Risk Level
SAFE
Analyzed
May 10, 2026, 12:36 PM
Security Audit — agent-trust-hub — us-program-research