Democratizing Dangerous Biology: How AI Lowers the Barrier to Biosecurity Risks

Author: Denis Avetisyan

New research reveals that artificial intelligence is making complex biological tasks accessible to a wider range of actors, raising concerns about the potential for misuse.

Leveraging large language model assistance significantly improves the performance of novice users on complex tasks, frequently achieving accuracy levels comparable to, and sometimes exceeding, those of both the language models themselves and seasoned human experts.

Large language models significantly improve performance on dual-use, in silico biology tasks, exceeding expectations and exposing gaps in current AI safeguards.

While increasing performance on biological benchmarks doesn’t necessarily translate to broader scientific advancement, our study, ‘LLM Novice Uplift on Dual-Use, In Silico Biology Tasks’, investigates whether large language models (LLMs) empower individuals with limited expertise to perform complex biological tasks. We found that LLM access yielded a substantial 4.16-fold increase in accuracy for novices-even surpassing expert performance on several benchmarks-raising concerns about both accelerated scientific progress and potential dual-use risks. Given that participants readily accessed potentially dangerous information despite safeguards, how can we effectively evaluate and mitigate the expanding accessibility of sophisticated biological knowledge through increasingly powerful AI tools?

The Ascendancy of Language Models in Biological Inquiry

Large Language Models are quickly establishing themselves as transformative instruments within biological research, driven by their capacity to navigate and consolidate vast datasets of complex information. These models, trained on extensive text and code, move beyond simple data retrieval, offering a novel approach to knowledge synthesis – identifying patterns, generating hypotheses, and even predicting outcomes with increasing accuracy. This capability is particularly valuable in biology, a field characterized by intricate relationships and an ever-growing body of literature, where extracting meaningful insights from sheer volume often presents a significant challenge. The potential extends to accelerating discovery across diverse areas, from drug development and genomic analysis to understanding disease mechanisms and predicting protein structures, effectively augmenting-and potentially reshaping-the landscape of biological inquiry.

Assessing the potential of Large Language Models in biological research demands more than anecdotal evidence; it requires standardized, challenging benchmarks. Recent work addresses this need by evaluating LLM performance on the Human Pathogen Capabilities Test (HPCT), a comprehensive assessment spanning core molecular biology and complex virology. Results indicate a substantial performance boost with LLM assistance, demonstrating up to a fourfold increase in accuracy compared to unaided analysis. This improvement suggests that these models aren’t simply recalling memorized facts, but actively contributing to more precise and effective biological reasoning, potentially accelerating discoveries in areas like pathogen identification and therapeutic development.

Restricting analysis to benchmarks with expert participation confirms that LLM-assisted novices (mean 27.0%, top 50% 50.3%) consistently outperform all other groups, demonstrating that the observed gains from human-AI collaboration are robust to variations in benchmark difficulty and composition.

Beyond Memorization: Benchmarks for True Biological Reasoning

Current biological reasoning benchmarks are designed to move beyond simple factual recall and instead evaluate an LLM’s ability to apply knowledge to practical tasks. These benchmarks include assessments of pathogen analysis, requiring models to interpret biological data and identify characteristics of disease-causing organisms. Furthermore, experimental design tasks are utilized to gauge a model’s capacity for formulating hypotheses, predicting outcomes, and outlining procedures for scientific investigation. The incorporation of these skill-based evaluations aims to determine whether LLMs can demonstrate genuine problem-solving capabilities within the domain of biology, rather than simply retrieving memorized information.

Current evaluations of Large Language Models (LLMs) extend beyond simple question-answering to assess their capacity for extended reasoning within complex biological problems. This is exemplified by ‘Long-Form Virology’ scenarios, which require LLMs to synthesize information and generate plausible designs for novel biological agents. These tasks are not focused on recalling existing virological data, but rather on the model’s ability to apply principles of virology to construct a coherent, albeit hypothetical, biological entity, demanding multi-step reasoning and the integration of diverse biological concepts to achieve a functional outcome.

The Agentic Bio-Capabilities Benchmark evaluates Large Language Model (LLM) performance in coding and problem-solving tasks situated within a biological context. Comparative analysis reveals that a ‘Treatment’ group, utilizing LLM assistance, demonstrated statistically significant improvement over a control group across nearly all benchmarked tasks. Notably, the Treatment group achieved performance exceeding that of human experts on several individual benchmarks, indicating LLMs are capable of not only assisting but, in specific instances, surpassing expert-level problem solving in this domain.

Analysis of participant performance on a virology benchmark reveals that the treatment group demonstrated both higher final scores and confidence levels compared to the control group, as evidenced by increased mean values ± standard error and positive correlations between submission time and performance.

Navigating the Dual-Use Dilemma: Responsible Innovation in Biology

Large language models are fundamentally reshaping biological research, yet this power introduces inherent ‘Dual-Use Biological Capabilities’. These models can accelerate discovery in areas like drug development and disease modeling, offering immense benefits to public health. However, the same capabilities that facilitate progress also present risks; the detailed biological knowledge generated can, in theory, be repurposed for harmful applications, including the creation of novel bioweapons or the enhancement of existing pathogens. This duality necessitates careful consideration of potential misuse, demanding proactive strategies to balance innovation with responsible development and deployment. It’s not simply about preventing malicious intent, but also about anticipating unintended consequences arising from the complex interplay between advanced AI and biological systems.

Effective risk mitigation for large language models in biological research hinges on the implementation of comprehensive biosafety protocols and continuous output monitoring. These protocols must extend beyond traditional laboratory safety measures to encompass the unique challenges posed by rapidly generated, computationally derived biological information. Ongoing monitoring isn’t simply about detecting obvious malicious intent; it requires sophisticated analysis to identify subtle inaccuracies, unintended consequences, or the potential for misuse even within seemingly benign applications. This proactive approach, guided by established ethical frameworks, aims to ensure that the benefits of LLM-driven biological innovation are realized responsibly, preventing the inadvertent creation or dissemination of information that could compromise public health or ecological stability.

A dependable benchmark of expert knowledge is paramount when evaluating large language models applied to biological research. These models, while powerful, can generate outputs containing inaccuracies or propose pathways with unforeseen consequences; therefore, establishing a curated ‘expert baseline’ – a dataset representing the consensus understanding of established biologists – allows for rigorous calibration of LLM performance. This comparison isn’t simply about identifying errors, but about discerning where a model excels, where it extrapolates beyond validated knowledge, and where its predictions deviate from established biological principles. Such a baseline enables researchers to quantify the reliability of LLM-generated hypotheses, assess the potential for unintended consequences, and ultimately, guide responsible innovation in a field where even subtle inaccuracies can have significant ramifications.

Participant responses reveal that LLM assistance was most valued for its direct support during tasks, whereas anticipated usefulness focused on broader conceptual aid.

Towards Reliable Outputs: Calibration and Troubleshooting of LLMs

Confidence calibration addresses the frequent misalignment between an LLM’s reported confidence score and the actual correctness of its output. Without calibration, a model may express high confidence in an incorrect response, leading users to inappropriately trust inaccurate information. Calibration techniques adjust the model’s output probabilities to better reflect empirical accuracy; a well-calibrated model assigning a 70% confidence to a statement will, over many similar statements, be correct approximately 70% of the time. This is achieved through post-hoc adjustments to predicted probabilities or, increasingly, by incorporating calibration loss functions directly into the model’s training process. Accurate confidence scores are critical for downstream applications requiring reliable decision-making and for identifying instances where human review is necessary.

Protocol troubleshooting within large language models (LLMs) is significantly enhanced by integration with sequence analysis and comprehensive literature review capabilities. This allows the LLM to not only identify discrepancies in biological protocols – such as reagent concentrations, incubation times, or procedural steps – but also to contextualize those errors. Sequence analysis facilitates verification of biomolecular sequences used in the protocol, while literature review provides access to established best practices and alternative methods. Combining these skills enables the LLM to suggest corrections grounded in both computational analysis and validated scientific literature, improving the reliability of complex biological workflows and reducing the potential for experimental errors.

Employing multiple Large Language Models (LLMs) in conjunction offers improved performance and reliability compared to single-model approaches, especially during prolonged interactions. This strategy capitalizes on the unique strengths of each model, creating a more robust system. Recent data indicates Gemini 2.5 Pro currently handles 33.5% of total user messages, demonstrating its significant contribution within a multi-model framework and highlighting the potential for distributed processing and increased system resilience.

Treatment participants demonstrate improved confidence calibration compared to controls, exhibiting less overconfidence-particularly at moderate to high confidence levels-as indicated by curves closer to the diagonal.

The study reveals a concerning reduction in the expertise required to engage with sophisticated in silico biology. This lowering of the barrier to entry, demonstrated by the ‘novice uplift’ observed with large language models, echoes a sentiment articulated by G. H. Hardy: “A mathematician, like a painter or a poet, is a maker of patterns.” While Hardy speaks of abstract patterns, the research highlights how LLMs now facilitate the creation of potentially dangerous biological patterns by individuals lacking traditional training. The elegance of a streamlined process, though seemingly benign, belies a risk – the pattern’s accessibility does not diminish its potential impact, and existing safeguards prove inadequate to control its proliferation. The core finding emphasizes a need for careful consideration of this accessibility, and a re-evaluation of existing biosecurity measures.

Where This Leads

The observed reduction in skill required to execute complex in silico biology represents a fundamental shift. The research does not prove malicious actors will inevitably emerge, only that the barrier to their entry has diminished. Current safeguards, assessed as insufficient, require re-evaluation – not through layered complexity, but through ruthless simplification. Focus should shift from detecting intent to limiting capability. The question is not “can harm be predicted?” but “can harm be prevented, given limited foresight?”

Benchmarking, presently fixated on performance metrics, must incorporate a measure of ‘novice uplift’-the degree to which a model empowers those lacking expertise. A model achieving state-of-the-art results is, paradoxically, less desirable if it simultaneously lowers the threshold for misuse. The field risks optimization towards increasingly powerful, yet less secure, systems.

Further research should investigate the interplay between model size, task specificity, and the resulting uplift. Clarity is the minimum viable kindness. The goal is not to halt progress, but to direct it towards systems demonstrably robust against unintended consequences – systems where simplicity, not sophistication, defines their ultimate safety.

Original article: https://arxiv.org/pdf/2602.23329.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

The Ascendancy of Language Models in Biological Inquiry

Beyond Memorization: Benchmarks for True Biological Reasoning

Navigating the Dual-Use Dilemma: Responsible Innovation in Biology

Towards Reliable Outputs: Calibration and Troubleshooting of LLMs

Where This Leads

See also: