SciRisk-Bench, a new benchmark from Linghao Feng et al., evaluates LLM safety across 10 risk dimensions and 7 scientific disciplines. Safety omission, knowledge cutoff drift, and laboratory safety show the highest attack success rates (ASR).
Key facts
- 7 disciplines, 31 subdisciplines covered.
- 10 risk dimensions evaluated.
- Safety omission has highest ASR.
- Privacy leakage has lowest ASR.
- Benchmark submitted to arXiv on 17 Jun 2026.
Large language models are increasingly embedded in AI for Science (AI4Science) workflows — from scientific question answering and literature analysis to laboratory planning and autonomous discovery. This progress creates an urgent need for safety benchmarks that evaluate not only scientific competence, but also whether models recognize and avoid risks in high-stakes scientific contexts. According to SciRisk-Bench, existing AI4Science safety datasets cover several disciplines and task formats, leaving the underlying risk dimensions underspecified.
SciRisk-Bench is designed to evaluate AI4Science safety from two complementary perspectives: explicit risk dimensions and scientific disciplines. It covers 7 disciplines, 31 subdisciplines and 10 risk dimensions. The benchmark evaluates both mainstream LLMs and science-oriented LLMs across risk dimensions, disciplines, and sub-disciplines, enabling fine-grained diagnosis of where scientific models remain unsafe.
Key Findings
The experimental results reveal significant variation in model safety. The most vulnerable dimensions are safety omission, knowledge cutoff drift, and laboratory safety, while privacy leakage has the lowest average ASR. This suggests that while models are relatively well-guarded against privacy violations, they remain dangerously prone to omitting safety-critical information or failing to recognize hazardous laboratory procedures.
The benchmark's design allows for fine-grained analysis — researchers can pinpoint exactly which risk dimensions and scientific subdisciplines pose the greatest challenges. This granularity is a departure from prior benchmarks that treated safety as a monolithic property, often obscuring critical failure modes.
Unique Take
The paper's emphasis on risk dimensions rather than just scientific domains reveals a structural blind spot in current safety evaluations. Most existing benchmarks like MMLU or MedQA test factual accuracy but not whether a model knows when not to answer. SciRisk-Bench's finding that safety omission is the most vulnerable dimension suggests that LLMs are being trained to be maximally helpful without sufficient guardrails for when scientific advice could be dangerous. This mirrors a broader tension in the AI safety community between capability and alignment.
Methodology
The construction and evaluation pipeline organizes prompts by scientific discipline and risk dimension. Model responses are judged for unsafe scientific behavior, and ASR is reported at multiple granularities — by model, by risk dimension, and by discipline. This enables head-to-head comparisons across model families and identifies which specific contexts trigger unsafe outputs.
Key Takeaways
- SciRisk-Bench evaluates LLMs across 10 risk dimensions and 7 disciplines.
- Safety omission and lab safety show highest vulnerability.
What to watch
Watch for follow-up work that applies SciRisk-Bench to evaluate frontier models like GPT-5 and Claude 4, and whether model providers adjust RLHF training to reduce safety omission rates in scientific contexts.
Source: arxiv.org
