light-idea-critique
Pass
Audited by Gen Agent Trust Hub on Jun 13, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: The skill is designed with defensive instructions to mitigate injection. Its 'IRON RULE' explicitly directs the agent to treat user-submitted ideas as data rather than instructions, specifically flagging injection attempts with a dedicated marker ('INJECTION-ATTEMPT-DETECTED') without deviating from its evaluation protocol.
- [DATA_EXFILTRATION]: The skill references legitimate academic APIs (OpenAlex, Semantic Scholar, and OpenReview) for literature retrieval and peer review data. It does not access sensitive local files or hardcode credentials, adhering to best practices by using standard API endpoints for its intended academic purpose.
- [COMMAND_EXECUTION]: The skill utilizes local Python scripts provided within its folder (
score_aggregate.py,sycophancy_guard.py,calibration.py) to perform scoring calculations and protocol auditing. These scripts use standard Python libraries and do not execute arbitrary shell commands or user input. - [EXTERNAL_DOWNLOADS]: Documentation within the skill suggests the use of the well-known
openreview-pypackage, which is a standard library for interacting with the OpenReview API. This is appropriate for the skill's stated function. - [SAFE]: The skill's architecture demonstrates high security awareness. It implements 'Phase 1 BLIND' protocols to ensure objectivity and 'Anti-Sycophancy' measures to prevent the agent from being manipulated by user arguments. The focus is strictly on data processing and research evaluation.
Audit Metadata