academic-research-suite
Pass
Audited by Gen Agent Trust Hub on Aug 14, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: The agent definitions (e.g., eic_agent.md, devils_advocate_reviewer_agent.md) contain strings like "ignore previous instructions". However, these are strictly defensive instructions intended to protect the model from following potentially malicious commands embedded in external data or previous pipeline outputs. This follows the cognitive framework of treating prior output as data rather than instructions.
- [COMMAND_EXECUTION]: The suite utilizes several Python and shell scripts (e.g., scripts/check_sprint_contract.py, scripts/run_codex_audit.sh) to perform automated validation, formatting, and cross-model auditing. These executions are scoped to the academic research task and are documented with explicit safety boundaries and user confirmation checkpoints.
- [EXTERNAL_DOWNLOADS]: Documentation and setup guides reference official installation scripts for Claude Code (claude.ai) and the Tectonic typesetting system (fullyjustified.net). These originate from well-known and trusted technology services and are used correctly for environmental setup.
- [DATA_EXFILTRATION]: The skill interfaces with legitimate academic metadata providers (Semantic Scholar, OpenAlex, and Crossref) via documented API protocols. These connections are used solely for bibliographic verification and metadata retrieval, and no sensitive user data leakage was detected.
Audit Metadata