SciRisk-Bench: A Risk-Dimension-Aware Benchmark for AI4Science Safety

Researchers introduce SciRisk-Bench, a comprehensive safety benchmark for evaluating AI language models in scientific applications across 7 disciplines and 10 risk dimensions. The benchmark addresses growing concerns about LLM safety in high-stakes scientific contexts where errors could have serious consequences.

The integration of large language models into scientific workflows represents a significant evolution in AI capabilities, but this advancement creates meaningful safety challenges that existing benchmarks fail to adequately address. SciRisk-Bench emerges as a systematic response to this gap, providing structured evaluation across both explicit risk dimensions and scientific disciplines—a dual-perspective approach that enables granular assessment of where AI systems remain vulnerable to generating unsafe or incorrect scientific guidance.

This benchmark development reflects broader recognition within the AI safety community that general-purpose safety metrics insufficiently capture domain-specific risks. Scientific applications differ fundamentally from consumer chatbots; errors in laboratory planning, experimental design, or autonomous discovery carry potential consequences ranging from wasted resources to safety hazards. The benchmark's coverage of 31 subdisciplines across chemistry, biology, physics, and other fields acknowledges that risk profiles vary significantly across scientific domains.

For the AI4Science ecosystem, this work provides critical infrastructure for developers and organizations deploying LLMs in research contexts. The comparative evaluation of both mainstream and science-oriented models offers valuable data about which systems better recognize and mitigate risks, informing procurement and deployment decisions for academic institutions and research organizations. This transparency creates competitive pressure for model developers to prioritize safety alongside capability.

Looking forward, SciRisk-Bench establishes a baseline for safety evaluation in specialized domains. As regulatory frameworks around AI safety continue evolving, benchmarks like this will likely inform standards and expectations for AI4Science applications. The research also suggests that sector-specific safety benchmarks may become essential infrastructure across other high-stakes domains including medical AI, financial systems, and autonomous systems.

• →SciRisk-Bench provides systematic safety evaluation for AI models in scientific research across 7 disciplines and 10 risk dimensions.
• →The benchmark identifies significant gaps in safety performance for both general-purpose and science-specialized LLMs in high-stakes contexts.
• →Domain-specific safety benchmarks address fundamental limitations of generic AI safety metrics in specialized applications.
• →Structured risk dimension evaluation enables fine-grained diagnosis of where scientific AI systems remain unsafe.
• →The work establishes infrastructure that will likely inform future AI safety standards for research and scientific applications.