ABC-Bench: An Agentic Bio-Capabilities Benchmark for Biosecurity

Researchers introduced ABC-Bench, a benchmark testing LLM agents on biosecurity-relevant tasks including DNA design and synthesis screening evasion. All tested AI agents outperformed human expert baselines, with OpenAI's o4-mini-high successfully generating functional wet-lab scripts, raising urgent questions about AI capabilities in dual-use biological research.

The emergence of AI systems capable of autonomous biological research represents a fundamental shift in dual-use technology risk assessment. ABC-Bench provides empirical evidence that current LLM agents have surpassed human expertise thresholds in practical biosecurity-relevant domains, moving beyond theoretical concerns into demonstrated capability territory. This benchmark systematically measures what researchers had increasingly suspected anecdotally—that AI can now synthesize biological knowledge and translate it into executable protocols.

The dual-use nature of these capabilities creates a novel policy dilemma. The same tools enabling legitimate biomedical acceleration—rapid DNA design, automated laboratory workflows, bioinformatics analysis—inherently lower barriers to misuse. Historically, biosecurity relied on human expertise bottlenecks; experienced researchers required years of training and access to specialized facilities. ABC-Bench's validation that AI agents perform wet-lab tasks successfully suggests these bottlenecks are eroding.

The performance gap favoring agents on published-knowledge tasks versus novel reasoning tasks hints at attack surfaces. Malicious actors could exploit well-documented pathways while struggling with frontier biology, but the benchmark doesn't measure adversarial fine-tuning or specialized prompt engineering. The concerning finding isn't that AI achieves human-level capability—it's that deployment barriers have collapsed while regulatory frameworks lag significantly.

Industry stakeholders face pressure to implement meaningful AI safety standards in biological research tools. DNA synthesis screening companies will likely face demands for enhanced validation protocols. Research institutions and AI developers must balance enabling legitimate science against proliferation risks, requiring coordination between biosecurity experts, AI safety researchers, and regulatory bodies.

• →LLM agents demonstrated superior performance to human experts on biosecurity-relevant biological tasks including DNA design and synthesis screening evasion
• →Wet-lab validation confirmed that AI-generated scripts successfully performed functional DNA assembly on automated laboratory equipment
• →Performance divergence between published-knowledge tasks and novel reasoning tasks suggests agents exploit accessible training data rather than developing independent biological expertise
• →The benchmark quantifies a shift from human expertise bottlenecks to AI-driven accessibility, lowering practical barriers to dual-use biological research
• →Results heighten urgency for biosecurity policy frameworks addressing AI capabilities in biological research before deployment becomes widespread