2 Sources
[1]
A defence-in-depth strategy for securing AI-driven biological sequence design
Generative artificial intelligence can design biological sequences, from proteins to whole genomes, opening opportunities for medicine and biotechnology, but the same capability can be misused. The Wübben Sandpit conference brought together experts in biology, AI, biosecurity, ethics, medicine and
[2]
AI in medicine and biology: Researchers call for a multilayered security architecture
Artificial intelligence is now capable not only of analyzing biological data but also of designing new DNA, RNA, and protein sequences. To lay the groundwork for responsible innovation, an international and interdisciplinary research team led by University Medical Center Mainz proposes a
Share
Copy Link
A team led from University Medical Center Mainz has published a plan for keeping generative AI in biological design safe, laying out layered safeguards across data, models and gene synthesis in a Nature Communications Perspective on 19 November 2026
1
2
.
Source: Medical Xpress
Their argument is that generative AI's dual-use potential grows with its capability, so governance has to evolve in step. Responsible governance, they argue, is a precondition for progress rather than a brake on it
1
. Per the paper, sequencing costs have fallen exponentially and produced a flood of data traditional analysis cannot keep up with, driving the adoption of generative AI in biology1
.Models already design functional proteins, enzymes and small viral genomes. A 2025 study fine-tuned a genome model on about 14,000 viral genomes and generated variants of the ΦX174 bacteriophage. After filtering, 16 of the 285 designs it synthesised worked as infectious phages
1
. The demonstrated capabilities stop at small, template-anchored viral genomes in benign lab systems. Designing entirely new, more dangerous human pathogens has not been shown; the concern is about what future capability could enable1
2
. The paper also names a structural gap: models can propose candidate sequences faster than any lab can screen or test them1
.The framework takes a defence-in-depth approach, with safeguards at three separate points
1
.On data, the paper recommends standardised classification of dual-use pathogen data, responsible release guidelines, and biosecurity review at the funding stage, with funders attaching risk-mitigation conditions to grants
1
. It also proposes tiered access on a five-level scheme, from open data up to material that could enable pandemic-capable variants1
. Filtering alone is not enough. Per cited biorisk evaluations, virus-relevant capability persists in open-weight genome models even when risky training data is removed, because researchers can reintroduce it during fine-tuning1
.On models, the recommendations cover lifecycle risk-benefit assessment, safety benchmarking that combines computational red-teaming with lab validation, and tiered deployment through hosted access
1
. Per the authors, no validated risk score for biological AI yet exists, so the model-side measures stay contingent until someone builds one1
.On synthesis, the final control point where a digital design becomes physical, the paper calls for stronger sequence screening that goes beyond simple similarity matching. It also wants identity checks and supply-chain tracking of benchtop synthesis devices, and a pooled insurance fund for the sector
1
. Screening is not mandatory in many countries, though many providers do it voluntarily1
. In software tests, screening could be evaded by AI-designed functional homologs, though a follow-up study found those homolog designs were not viable in wet-lab conditions as safe proxies1
.Workshop participants judged the upstream levers, funding conditions and data-sharing agreements, the most effective instruments, because they act before a dangerous sequence reaches a synthesiser
1
.Related Stories
The proposals came out of a three-day workshop in Germany in November 2025, funded by the Wübben Wissenschaftsstiftung. Its twenty participants came from molecular biology, AI, biosecurity, ethics, medicine and policy
1
2
. No formal consensus vote was taken; the authors consolidated the directions afterwards1
.They also list the risks of their own proposals. Restrictions could be too broad or too narrow, or split research into a two-tier community that disadvantages under-resourced groups and the Global South. Well-funded malicious actors might simply build clandestine copies of restricted datasets, potentially handing attackers an edge over defenders
1
. The measures would apply to future data only, and dual-use pathogen data is a small proportion of all biological data1
.There is a live policy vehicle waiting. The proposed EU Biotech Act contains closely aligned provisions, from customer verification to mandatory screening on benchtop synthesis devices. It frames biosecurity in terms of competitiveness and strategic autonomy, presenting security-by-design as a European selling point
1
.Lead author Dr. Maximilian Sprang, a dermatology researcher who leads a group at the Institute for Quantitative and Computational Biosciences at Johannes Gutenberg University Mainz, warns against blanket restrictions. "It is crucial to realistically assess what AI is already capable of and to understand its risks early on," he said.
2
Summarized by
Navi
29 Apr 2025•Science and Research

03 Oct 2025•Science and Research

05 Jun 2026•Policy and Regulation
