## The problem

One of the most fortunate facts about the current biosecurity landscape: there are essentially no credibly identified pandemic-potential pathogens beyond smallpox, and even smallpox isn't catastrophically bad. This means that even if our other defenses fail, bad actors struggle to identify catastrophic targets.

This buffer is closing.

- Recent advances in biological design tools (BDTs) democratize the capability to create dangerous novel pathogens. [These tools are often released open-source with effectively no guardrails](https://www.rand.org/randeurope/research/projects/2024/ai-risk-index.html).
- General-purpose LLMs may suggest dangerous pathogens to bad actors and make attacks more feasible.
- Pathogen functional characterization data, particularly Deep Mutational Scanning data on immune evasion or receptor binding, is a plausible chokepoint. These datasets are currently rare, expensive, and technically challenging to generate, which makes them a promising control point.
- The window is closing fast. As wetlab automation improves and more labs publish openly, the chokepoint erodes.

Without empirical work on whether controls actually delay capabilities, and without the policy scaffolding to implement them, this chokepoint will close before anyone has tried to use it.

## What winning looks like

Within 18-24 months:

- A new organization or housed project has answered the core empirical questions about biological data controls.
- Concrete policy proposals (tiered Biosafety Data Levels, Trusted Research Environment standards, grant conditionality, export controls) are on the table at HHS, NIST, and federal funding agencies.
- At least one data classification tier is enacted at federal level, ideally with international coordination underway.

## Why now

BDTs and AI models are advancing on a fast clock. The federal AI safety and biosecurity policy environment is unusually open to new instruments. The AI safety and biosecurity policy communities are both maturing; collaboration is timely. The chokepoint window closes more each year that nobody acts.

## What this could look like in practice

- **Independent research nonprofit:** Empirical ML work plus policy translation, founded specifically for this purpose.
- **Project inside an existing think tank:** Housed at RAND, CLTR, or a similar org with existing policy reach.
- **Policy advocacy org:** Focused on the legislative and executive policy push, partnering for empirical work rather than running it in-house.

## Who should do this

- **Profile:** ML research background plus biology comfort (or vice versa). Or a policy expert with deep interest in AI-bio convergence and willingness to commission and interpret empirical work.
- **Archetype:** Researcher or Policy person.
- **Commitment:** full-time, 6+ months minimum. Realistically 2+ years to a meaningful policy outcome.

## How to start

1. Identify in silico proxies for dangerous biodesign capabilities and design the first experiment: measure whether adding or removing a candidate dataset measurably shifts model performance on those proxies.
2. Map the publishing pipeline for the highest-risk data types (Deep Mutational Scanning, etc.) by country and lab. Estimate the share of future data that would fall under compliant jurisdictions.
3. Draft a tiered classification proposal (Biosafety Data Levels) and circulate to a small group of policy stakeholders for input.

## Source and endorsements

- **Field strategy:** BlueDot AI-bio convergence field strategy. (Link TBD.)
- **Relevant prior work:** [Science article on dangerous biodesign capabilities](https://www.science.org/doi/10.1126/science.aeb2689), [supporting policy doc](https://docs.google.com/document/u/2/d/1BRFCNW65O8wbEjuhoSc1Hv2ZDk-2cjhKgOXHFj2svXA/edit?usp=sharing).
- **Partners and collaborators:** Saif Khan, Luca Righetti, CHS, Biohub, RAND, CLTR.
- **Currently pursued by:** Not yet.

## Get involved

- **Funding pathway:** BlueDot exploration grant up to $20K, career transition grant up to $60K. Larger raises (~$500K) via Coefficient Giving for the founding sprint.
- **Advisors and introductions:** BlueDot can connect you to Saif Khan, Luca Righetti, and contacts at CHS, Biohub, RAND, and CLTR.
- **Contact:** dewi@bluedot.org


Pandemic-potential pathogens are still rare and hard to identify, which buys us safety even if our other defenses fail. Biological design tools and AI models are eroding that buffer. Controlling access to the narrow category of training data that enables misuse-relevant biodesign capabilities could buy meaningful time, but nobody is doing the empirical work to confirm it would, or building the policy infrastructure to implement it.

Build biological data access controls to delay the development of high-risk biodesign capabilities