Agentic Bio-Capabilities Benchmark Released for AI Laboratory and Biosecurity Evaluation

Researchers introduced the Agentic Bio-Capabilities Benchmark (ABC-Bench) in a study, Sciencecast reported. The benchmark comprises a suite of tasks that require large language models to program liquid-handling robots, design DNA fragments, and evade DNA synthesis screening. When evaluated, the AI agents outperformed average human experts on these challenges.

The results also highlighted the dual-use nature of the technology, raising concerns about potential biosecurity risks. The release provides a standardized tool for measuring AI capabilities in laboratory and bioinformatics contexts.