A new paper, Williams et al. 2026 is out. It provides better-grounded evidence that current rapid ongoing increases in AI capability make it easier for people to perform the sophisticated work needed to modify pathogenic viruses and cause harm. URL and pointer to explainer in comments.
Biosecurity Really, Drew Endy et al.'s report on securing the biological world, has been out for a couple weeks. URL follows. It's full of true facts and everybody should read it and harken to its policy recommendations. The tough recc I'll single out here is prohibition of some virological research, both to lower danger and to start establishing a measure of public and international trust.
Consequences. 1) For this approach to operate at all, model developers would need to blast out so big a hole in the training data as to create real demand for safety-detuned open weight models trained on entire literature 2) model developers will be surprised how well today's and tomorrow's models will be able to reason and infer their way from the edges of the hole back towards its center.
Chen et al. Anthropic, August 19, 2025. Enhancing Model Safety through Pretraining Data Filtering. With respect for the authors, two of whom I know and like and another whose work I know and admire. For biology, this approach can't work. Not without hampering scientists' ability to use models to do science.
Two reasons. 1) well intended work on pathogens will always be dual use. 2) (less widely appreciated). 2) Mechanisms and genetic elements of pathogens have been pressed into service throughout biological R&D and there is ever growing use of virus-like and virus-inspired mechanisms and contraptions in in medicine
Consequences. 1) For this approach model developers would need to blast out so big a hole in the training data as to create real demand for safety detuned open weight models trained on entire literature 2) model developers will be surprised how well today's and tomorrow's models will be able to reason their way from the edges of a hole back towards its center.
Consequences. 1) For this approach model developers would need to blast out so big a hole in the training data as to create real demand for safety detuned open weight models trained on entire literature 2) model developers may be surprised how well today's and tomorrow's models will be able to reason their way from the edges of a hole back towards its center
Two reasons. 1) well intended work on pathogens will always be dual use. 2) (less widely appreciated). 2) Mechanisms and genetic elements of pathogens have been pressed into service throughout biological R&D and ever growing use of virus-like and virus-inspired mechanisms and contraptions in in medicine.
@AnthropicAI With respect for the authors, two of whom I know and like and another whose work I know and admire. For biology, this approach cannot work. Ever. Not without hampering scientists' ability to use models to do science.
This is to praise Susan Coller Monarez, @CDCMonarez. Monarez a textbook example of scientist as public servant and the good such scientists can do. Now forced out of government. But her continuing work will surely help defend the US and the human population in years ahead.
Next post links to a short review of Kadlec report. First thing to mention is that, if one believes that heroes exist, then people who by their brains and actions save millions of lives, such as Kadlec, then ASPR at HHS, and Dr. Peter Marks, then head of CBER at the FDA, qualify as heroes. Pure and simple. Next. Kadlec's report deals not with what historians will regard as causation, but with the narrower and unfortunately more contested issue of how SARS-CoV-2 entered the human population. This is not the same question as the question of what caused the pandemic. If there is to be a human history, that larger question will be the subject of work by historians for decades to come
@davidmanheim David, Shakeel, @ShakeelHashim, isn't it beyond worrisome that the EU used the term "systemic risks" incorrectly? With such breezy bureaucratic confidence? Given what we believe to be the stakes? And the need to get things right the first time?
This partnership is the perfect setup for Anthropic to demonstrate the most promising way to deal with biorisks:
1. Deploy a general-purpose LLM with its virology & bio-knowledge nuked (filtered/unlearned etc.)
2. Deploy on a specialized high KYC platform the version with bio