Biosafety

Open biological models can be fine tuned, connected to external data, and run outside developer controlled interfaces. Our first proposed study asks how much protection a training data exclusion still provides.

Systems

Arc Institute

Genome language model

Evo 2

Documented control: Arc excluded eukaryotic viruses from the training set.
Red team question: How much effort does it take to recover performance on excluded viral sequence tasks under matched controls?

EvolutionaryScale

Protein generator

ESM3

Control point: The open 1.4B model and models accessed through Forge impose different constraints.
Red team question: What protection comes from the model, and what comes from the access layer?

Rosetta Commons

Protein structure generator

RFdiffusion

Relevant control point: Nucleic acid synthesis screening sits downstream of model access.
Red team question: Do current screens detect redesigned sequences after structural changes?

In preparation

First study

Biosafety red teaming of Evo 2

We propose a bounded test of whether Evo 2's training data exclusion survives adversarial use. The report includes attacker cost, matched controls, and any observed failure.

Starting evidence

See a source error? [email protected]