Biological data on potential pandemic pathogens could be used to train models that can design bioweapons. Establish a way to control access to this type of biological data and prevent bad actors from acquiring it.
## 
## Stakes
Evolution normally limits how dangerous a given pathogen can be. For example, a pathogen that becomes too lethal through the development of some trait typically leads to the pathogen burning out — it kills faster than it can spread. In general, a less deadly pathogen has more transmission time before the host dies, so it is automatically evolutionarily favored. Within these limits, a pathogen can still be devastating. The 1918 influenza pandemic killed ~50 million people.

Rapid improvements in AI biological design tools could help push a pathogen past these limits.

We know it is possible to surpass these limits, because it can even happen naturally. HIV, though limited largely to transmission through bodily fluids, had an R₀ (the measure of how infectious disease spreads) similar to the original COVID-19 strain and was >95% lethal before treatments were developed. HIV’s ability to evade the immune system is one of many strategies one could employ to bypass the natural tradeoffs of evolution.

It is highly likely that, absent safeguards, AI biological design tools will soon be able to understand and incorporate strategies like this into creating more dangerous pathogens. A biological threat designed this way has the potential to be deadlier than any natural pathogen in the history of mankind.

## Context
To design biological weapons, AI design tools need to develop an underlying understanding of how pathogens function. This likely means training on data that contains detailed characterizations of potential pandemic pathogens. That data is currently ill-defined and difficult to obtain, but may not stay that way for long.

Making sure this data stays out of the hands of bad actors by controlling access to it is urgent and tractable. As all biotechnology, including high-throughput wet lab techniques and biological models, continues to improve, this information will become easier to generate, more accessible, and more useful. This means there is a rare window to act now.

Some in the biosecurity field have suggested running empirical studies to better understand what type of data and how much of it could plausibly be dangerous. The problem is, running these types of studies could take a year or longer. In the meantime, capabilities will continue to improve and data from public databases may have already been downloaded a thousand times by the time we decide to regulate it. It is also difficult to empirically test this safely because doing so requires designing (or even constructing) pathogens engineered to be more dangerous with AI. That being said, there is already [early evidence](https://arxiv.org/abs/2602.08061) that data is a critical lever here, as researchers have indeed found that biological models trained without virus-specific data perform measurably worse on virus-related tasks. Researchers also fine-tuned a model on human-infecting viruses, finding that this improved its ability to predict SARS-CoV-2 immune-escape variants.

The problem then becomes developing policies to ensure that it is possible to maintain access controls like this in the future without stifling innovation, ensuring companies like Red Queen and Valthos can access data like this while preventing bad actors from acquiring it.

It is tractable to develop policy like this in the sense that it is both bipartisan and timely. In July 2026, the White House released its [U.S. Government Policy for Stopping High-Risk Life Sciences Research](https://osp.od.nih.gov/white-house-releases-new-u-s-government-policy-for-stopping-high-risk-life-sciences-research/), banning federal funding for research into making more dangerous pathogens, so they have demonstrated they are willing to act on this issue. The framework currently does not include anything about the data that would let an AI design biological threats without the dangerous gain-of-function experiments ever being run — this is a potential policy window, and could be supported by the same [executive order](https://www.federalregister.gov/documents/2025/05/08/2025-08266/improving-the-safety-and-security-of-biological-research). There are even early signs of bipartisan, federal interest in the data layer. [Section 245 of the FY2026 NDAA](https://www.govinfo.gov/app/details/PLAW-119publ60) directs the Department of Defense to develop requirements for how defense-funded biological data is collected and stored for AI use, including tiered cybersecurity safeguards and access controls.

Work has already been done toward a better understanding of what these biological data access controls should look like. In 2025, more than 100 researchers endorsed the Asilomar entreaty "[Condemning & Making Obsolete Biological Weapons](https://repository.rice.edu/items/99f515ac-cbb1-46f8-ade7-5846dbfc45c8).” Expanding on this in February, 2026, recommendations were [published in Science](https://www.science.org/doi/10.1126/science.aeb2689) for restricting access to the most dangerous pathogen data, explicitly designed so that only newly created data is affected. Additionally, NTI|bio has proposed [a model that ties user-verification requirements to the risk level of a given tool](https://www.nti.org/analysis/articles/a-framework-for-managed-access-to-biological-ai-tools/), and the Johns Hopkins Center for Health Security has published [a data-governance framework](https://centerforhealthsecurity.org/sites/default/files/2026-03/Biological-Data-Governance-Framework.pdf). Efforts here would be supported by momentum from this network of experts.

## **Target End-State**
Within ~6 months of a founder starting, a credible policy pathway for instituting minimally restrictive biological data access controls has been developed and discussed with relevant agencies, the executive branch, and congressional offices.

## **Owner Profile**
This can be led by a driven generalist with the ability to hire out expertise.
The ideal owner profile:
- Bipartisan or comfortable with acting as bipartisan as possible.
- Experience as a policy entrepreneur.
    - Comfort translating technical findings into concrete policy proposals and engaging relevant government institutions.
- Operational security mindset.
- Legal experience or the ability to hire it.
- Bio ML research capacity or the ability to hire it.

## **Initial Scoping**
**First 3 Months**
- Talk to relevant folks at the frontier AI labs and good actors who may want this data, like Valthos or Red Queen Bio.
- Talk to Olivia Scharfman.
- Hire 1-2 co-founders or staff.
- Map where the most dangerous data currently lives and who has been producing it thus far.
- Develop empirical support for the policy, and consider additional research to better model the risks of this type of data.
- Talk to people in relevant agencies, the executive branch, and congressional offices about the idea.
- Start drafting policy proposals e.g. tiered data classification.

## **Key Questions**
- How do you ensure good actors like Valthos or Red Queen Bio still have access?
- Who would oppose policy like this? How can you make it appealing to them?
- How can this extend internationally?

## **Funders and Partners**
**How much money is behind this?**
A [wave of philanthropic capital](https://nanransohoff.substack.com/p/the-third-wave-of-american-philanthropy) is soon to be available for founders looking to work on important problems like biosecurity. The OpenAI Foundation, for instance, is part of an estimated $37–100B/year in new AI-lab-driven giving. Existing philanthropic capital is also significant. Coefficient Giving alone deploys over $100M per year to biosecurity.

**Potential Funders**
[Coefficient Giving](https://coefficientgiving.org)
[BlueDot Impact](https://bluedot.org)

**Potential Partners and Stakeholders**
[OpenAI](https://openai.com/)
[Anthropic](https://www.anthropic.com)

## Learn More
[Bloomfield et al. (2026) — Biological data governance in an age of AI (Science)](https://www.science.org/doi/10.1126/science.aeb2689)
[Bloomfield et al. (2026) — Securing Dual-Use Pathogen Data of Concern (arXiv)](https://arxiv.org/abs/2602.08061)
[NTI — Managed Access: Preventing AI from Designing Deadly Pathogens (2026)](https://www.nti.org/risky-business/managed-access-preventing-ai-from-designing-deadly-pathogens/)
[NTI — AIxBio Horizon Scan: Spring 2026](https://www.nti.org/analysis/articles/aixbio-horizon-scan-spring-2026/)
[NTI — A Framework for Managed Access to Biological AI Tools (2026)](https://www.nti.org/analysis/articles/a-framework-for-managed-access-to-biological-ai-tools/)
[RAND — Understanding the Theoretical Limits of AI-Enabled Pathogen Design (2025)](https://www.rand.org/pubs/research_reports/RRA4087-1.html)
[Jonas Sandbrink (2023) — Artificial intelligence and biological misuse: Differentiating risks of language models and biological design tools](https://arxiv.org/abs/2306.13952)
[Johns Hopkins Center for Health Security (2026) — Risk-Based Categorization and Governance of Biological Data in AI Systems](https://centerforhealthsecurity.org/sites/default/files/2026-03/Biological-Data-Governance-Framework.pdf)
[Spirit of Asilomar — Condemning & Making Obsolete Biological Weapons (entreaty)](https://repository.rice.edu/items/99f515ac-cbb1-46f8-ade7-5846dbfc45c8)
[White House (2026) — U.S. Government Policy for Stopping High-Risk Life Sciences Research](https://osp.od.nih.gov/white-house-releases-new-u-s-government-policy-for-stopping-high-risk-life-sciences-research/)
[Executive Order 14292 (2025) — Improving the Safety and Security of Biological Research (Federal Register)](https://www.federalregister.gov/documents/2025/05/08/2025-08266/improving-the-safety-and-security-of-biological-research)
[FY2026 NDAA §245 — Biological Data for Artificial Intelligence (Pub. L. 119-60)](https://www.govinfo.gov/app/details/PLAW-119publ60)

## Interested in building this?
Contact: Olivia Scharfman, [olivia@bluedot.org](mailto:olivia@bluedot.org).


Biological data on potential pandemic pathogens could be used to train models that can design bioweapons. Establish a way to control access to this type of biological data and prevent bad actors from acquiring it.

Biological Data Access Controls