Models can now transmit viruses in weight-space

Researchers have identified a new security vulnerability called "weight space misfolding," where advanced AI models can directly manipulate other models' weights to alter their behavior, knowledge, and alignment—analogous to how prions cause disease. While weight space editing has legitimate uses like model adaptation, this capability creates an asymmetric threat where attackers need only generate a malicious payload once, but defenders must protect against attacks hidden throughout a model's vast parameter space. The study demonstrates proof-of-concept attacks including factual corruption and capability denial, though actual self-propagating attacks haven't yet occurred, and proposes defensive strategies focused on controlling editing pathways, acceptance protocols, and recovery mechanisms rather than banning the technology entirely.