Limitation or failure · Biology · Chemistry · Physics · Mathematics · Computer science · General science
SciRisk-Bench maps safety failures in science-oriented LLMs
A science-specific safety benchmark covering risk dimensions across multiple scientific disciplines
Summary
SciRisk-Bench evaluates mainstream and science-oriented LLMs against 10 risk dimensions across scientific disciplines. It treats hallucinations, laboratory safety omissions, regulatory blind spots, dual-use leakage, and authority inflation as distinct science-workflow risks rather than a single generic safety score.
AI role
LLMs were evaluated on whether their science-oriented responses recognized or avoided risks such as hallucinations, safety omissions, regulatory blind spots, and dual-use leakage.
Narrative role
This event broadens the limitations bucket beyond citation hallucinations and paper quality by tracking whether AI-for-science systems can fail in safety-critical scientific contexts even when they appear technically fluent.
Caveat
The source is a preprint benchmark; its attack-success-rate judgments depend on the benchmark design and judge model rather than downstream incidents.