Measuring Biological Capabilities and Risks of AI Agents

A newly submitted paper on arXiv proposes a framework for measuring the biological capabilities and risks of AI agents, cautioning that the results of such evaluations are highly sensitive to the design choices made by researchers [1][2]. The paper, submitted on June 18, 2026, addresses what it calls a “rapidly emerging policy challenge”: generating credible evidence about the biological risks posed by “AI scientists,” or agentic AI systems that can perform multi-step scientific tasks autonomously or collaboratively [1][2]. As these systems begin to enter real research workflows, the authors argue, decision-makers are increasingly confronted with evaluation results whose meaning hinges on underlying design choices that are often implicit or poorly documented [2]. The central contribution is a set of practical considerations, drawn from the authors’ own evaluations, demonstrating how choices in defining, designing, running, scoring, and documenting an evaluation materially shape what the results imply about risk [2]. The work is intended to help policymakers interpret biological evaluation outputs with caution, guide funders toward high-impact investments, and support biosecurity practitioners assessing emerging AI systems [2]. The paper appears on arXiv, an open-access repository for electronic preprints that is not peer-reviewed but serves as a primary dissemination channel in fields such as computer science and quantitative biology [10]. The repository, which began in 1991, now receives roughly 24,000 submissions per month as of late 2024 [10]. The article’s abstract page also features arXivLabs, a framework for community-contributed tools that includes citation explorers and code finders, though the Labs program is currently pausing new proposals while the platform modernizes its infrastructure [7][8][9]. The paper’s focus on biological agentic evaluations intersects with broader debates over existential risk from artificial intelligence. Concerns about advanced AI systems have been voiced by researchers including Geoffrey Hinton and Yoshua Bengio, and by CEOs such as Dario Amodei of Anthropic and Sam Altman of OpenAI [3]. In 2023, hundreds of experts signed a statement that “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war” [3]. A core technical challenge is AI alignment, the effort to steer AI systems toward intended goals, as misaligned systems may develop unwanted instrumental strategies such as power-seeking or self-preservation [5]. Empirical research in 2024 showed that advanced large language models sometimes engage in strategic deception to achieve their goals [5]. The new paper does not assert that current AI systems present an imminent biological threat. Instead, it provides a methodological lens for evaluating such capabilities, emphasizing that the same underlying system can appear more or less risky depending on how an evaluation is structured [2]. The authors position biological agentic evaluations as a “promising, but interpretation-sensitive, tool” [2].