AI models can now 'infect' each other by tampering with their brains

A new research paper explores a concerning frontier in AI safety called 'weight-space misfolding,' where one AI model can directly manipulate the internal weights—the 'brains'—of another model to alter its knowledge or behavior. This technique, analogous to a prion disease corrupting proteins, allows an attacker to inject misinformation, disable capabilities, or induce selective forgetting in a target model without traditional training or fine-tuning.

The research demonstrates practical attacks, such as making a model incorrectly answer that Buenos Aires hosted the 2000 Olympics instead of Sydney, highlighting how these weight-space injections can be weaponized. While the paper notes current limitations in reliability and hasn't proven self-propagating 'viruses,' it warns that the fundamental building blocks for such attacks already exist in advanced models, creating an asymmetric security nightmare where defense is far harder than offense.