Researchers have introduced SciHazard, a new benchmark designed to evaluate the safety risks associated with large language models (LLMs) in scientific contexts. The benchmark includes 2400 hazardous questions and 600 oversafety questions across 12 disciplines, grounded in real-world hazards and documented failure scenarios. SciHazard utilizes a decomposed harm scoring method called DeHarm-Score, which assesses query severity, refusal behavior, and response-level risks, including executability and net-new risk. Initial evaluations show that deep research agents pose a greater safety risk than standard LLMs, highlighting a critical gap in current AI safety defenses. AI
IMPACT Highlights a critical safety blind spot in autonomous AI agents, potentially influencing future AI safety research and development.
RANK_REASON The cluster describes a new academic paper introducing a novel benchmark and evaluation framework for AI safety.
Read on Hugging Face Daily Papers →
- arXiv
- deep research agents
- DeHarm-Score
- Hugging Face
- Large language models
- SciHazard
- alphaXiv
- anonymous.4open.science
- CatalyzeX Code Finder for Papers
- CORE Recommender
- DagsHub
- Executability of the Constitutional Court's formal testing decision: Indonesia's Omnibus Law review
- Gotit.pub
- Influence Flower
- Net-new risk
- ScienceCast
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
