Biosecurity Experts Warn of Risks from AI-Designed Viral Genomes in Science Commentary

In a perspective published in Science, biosecurity researchers Thomas V. Inglesby and Moritz S. Hanke detail the biosafety, biosecurity, and governance implications of generative artificial intelligence models capable of designing functional viral genomes. The commentary addresses the dual-use risks and policy gaps highlighted by recent breakthroughs in genome language models that construct viable synthetic bacteriophages from scratch.

Generative Genome Design and Biosafety Implications

The publication in Science by Inglesby and Hanke, titled “AI-designed viral genomes,” examines the emergence of generative AI architectures capable of composing complete, functional viral DNA sequences. Rather than processing natural language, genome foundation models like Evo process raw nucleotide sequences at single-nucleotide resolution across billions of base pairs.

While the technology offers applications for therapeutic development, the authors emphasize that creating fully synthetic viral genomes presents severe biosecurity challenges. Inglesby and Hanke note that while the ability to compose viral genomes using generative AI now exists, the regulatory and safety governance required to steer the technology safely does not.

The commentary points out key biosecurity considerations raised by generative biology:

• Dual-Use Capabilities: Generative models trained on biological sequences can design functional biological agents, raising concerns that the same computational tools used for medicine could be misapplied.
• Countermeasure Evasion: Synthetic viral genomes generated by AI may encode novel sequence patterns that bypass natural immune recognition, diagnostic tools, or existing therapeutic interventions.
• Governance Gaps: Existing oversight frameworks primarily focus on regulated physical pathogens rather than digital, open-weight generative models capable of generating novel sequence designs.

Generative DNA Sequence Design Pipeline:

[Training Corpus] ----> [Genome Language Model] ----> [Candidate Genomes]
(Non-pathogenic)        (Nucleotide Resolution)       (In Silico Generation)
                                                                |
                                                                v
[Host Cell Infection] <--- [In Vitro DNA Synthesis] <--- [Filtered Candidates]
(Biological Replication)   (Lab Assembly)                (Selected Sequences)

Experimental Validation of Synthetic Bacteriophages

The commentary addresses empirical findings from accompanying research demonstrating the physical viability of AI-generated viruses. Researchers utilized the Evo genome language model to generate thousands of prospective viral sequences modeled after the Phi X-174 bacteriophage, a small virus that infects bacteria.

The experimental workflow progressed through structured computational and laboratory stages:

1. The genome language model evaluated sequence patterns across millions of natural bacteriophages to generate over 700,000 candidate sequence variations.
2. In silico filtering algorithms selected approximately 300 candidate genomes for physical laboratory synthesis.
3. Synthetic DNA sequences were synthesized and introduced into Escherichia coli host cells.
4. Host cellular machinery transcribed and translated the synthetic genetic code, producing assembled viral particles.
5. Assays confirmed that 16 of the AI-designed bacteriophage genomes were biologically viable, successfully infecting, lysing, and replicating within E. coli bacterial strains.

In laboratory testing, a cocktail composed of these 16 synthetic bacteriophages successfully killed strains of E. coli that had developed resistance to natural, wild-type bacteriophages. This result demonstrates the potential for generative AI to assist in creating adaptive phage therapies to target antibiotic-resistant bacteria.

Model Training Safeguards and Dual-Use Risks

To prevent the creation of dangerous biological agents during initial experiments, researchers implemented explicit training safeguards. Pathogenic viruses and organisms capable of infecting humans, animals, or complex plants were excluded from the model training corpus. Additionally, physical laboratory testing was restricted to non-pathogenic bacterial host strains within controlled containment facilities.

However, Inglesby and Hanke warn that similar model architectures could be trained on datasets containing human, animal, or plant pathogens. Training generative models on dangerous viral families could enable the automated design of novel pathogens with enhanced transmissibility, altered host tropism, or increased virulence.

The commentary highlights specific technical characteristics of the model training methodology:

• Training Corpus Scope: The primary training dataset comprised billions of base pairs from bacterial and phage single-cell organisms while deliberately omitting human viral pathogens.
• Model Scale: Advanced variants of the Evo foundation model incorporate up to 20 billion parameters, permitting long-context context windows capable of modeling whole viral genomes.
• Resolution: Sequence generation operates directly on individual A, C, G, and T nucleotide bases without relying on pre-segmented gene annotations or natural language prompts.

Governance Frameworks and Policy Recommendations

Inglesby and Hanke argue that research aimed at generating novel viral pathogens capable of infecting humans, animals, or plants should not be pursued due to catastrophic risk potential. They call for international consensus on screening protocols for synthetic DNA orders and stricter oversight of open-weights generative models in the life sciences.

Existing regulatory measures, including the United States Government Policy for Stopping High-Risk Life Sciences Research, target dangerous gain-of-function research involving known high-consequence pathogens. However, the authors argue that generative biological AI expands the threat surface beyond known agent lists to include entirely novel synthetic sequences.

The commentary concludes that biosecurity governance must adapt to manage open-access generative biological software, requiring mandatory biosecurity reviews, strict DNA synthesis provider screening, and international risk assessment standards prior to model deployment.