About This Report  iii

Summary  v

Figures and Tables  ix

Interpreting Dual-Use Biology Benchmarks for Frontier AI; Measurement, Saturation, and Generational Analysis  1

Introduction  1

Validity  2

Contamination and Refusal  3

Structure of the Report  3

Methods  4

Item Response Theory  5

Results  5

Summary Results  6

Dataset Composition  9

SecureBio Dual-Use Biology Benchmarks with Expert Baselines  11

Interpreting Frontier Performance  15

Saturation  15

More Efficient Assessments  17

Generational Analysis of Gemini  18

Discussion  20

A Theoretical Framework for Interpreting Capability-Relevant Constructs  20

Theories of Human Uplift  20

Limitations  22

Conclusion  23