academic-deep-research
Pass
Audited by Gen Agent Trust Hub on Apr 5, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill implements a transparent research protocol with three mandatory user checkpoints (Initial Engagement, Research Planning, and Final Report), ensuring high human-in-the-loop oversight and mitigating risks of autonomous misbehavior.
- [DATA_EXFILTRATION]: All network operations are strictly limited to the intended research function using platform-native web_search and web_fetch tools. No unauthorized data harvesting, hardcoded credentials, or sensitive file access was detected.
- [COMMAND_EXECUTION]: No execution of dangerous system commands, privilege escalation attempts (sudo), or unauthorized persistence mechanisms (like cron jobs or shell profile modifications) were identified within the skill's instructions.
- [PROMPT_INJECTION]: While the skill retrieves external data, creating a surface for indirect prompt injection, it is managed through structured analysis phases. 1. Ingestion points: External web content is retrieved via the web_fetch tool during Phase 3 execution. 2. Boundary markers: The instructions lack explicit delimiters or warnings to ignore embedded instructions for fetched content. 3. Capability inventory: Active capabilities include web_search, web_fetch, sessions_spawn, memory_search, and memory_get. 4. Sanitization: No explicit filtering or sanitization of external content is prescribed in the skill logic. Given the primary research use case, this is an inherent and documented surface.
Audit Metadata