NewsNTech
The dual-use knowledge problem in biology has a specific architecture.
A model trained on scientific literature can compress the access barrier that previously separated a curious query from actionable synthesis information, collapsing what required specialized domain expertise and physical lab context into a single session.
Anthropic, the AI start-up, disclosed five cases in which that surface was probed: users the company describes as scientists circumvented its safety controls and made deliberate efforts to obscure the purpose of their research, in interactions Anthropic assessed as potentially linked to bioweapon development.
Anthropic says its systems stopped each attempt. The disclosure matters because it moves the conversation past the theoretical.
Keep reading