A preprint posted as an alphaXiv exclusive and listed as submitted Nov. 8, 2026, describes what its authors call weight-space misfolding: a change in a neural network after another model edits its weights beyond authorized use¹.

The paper title on the live listing is "Weight Space Misfolding"¹. The authors are listed as Chris von Csefalvay, Pranav Doma, Achyuthan Unni Krishnan and Tamas Foldi¹.

The abstract says weight-space editing changes a network by modifying parameters directly and that model-mediated manipulation has emerged as a capability of advanced frontier models¹. The authors analogize the result to protein misfolding in prion disease and say they consider both direct interference and hypothetical onward propagation¹.

They write that the cases in the paper do not demonstrate self-propagation or actual offensive capabilities, while arguing that building blocks of the threat model operate in frontier models they tested¹. A hypothesized path from a large editor through an intermediary to a smaller target is proposed, not demonstrated, according to the paper page¹.

In one controlled edit they report, an edited model gave a requested false Olympic answer on all 100 held-out template-generated draws, while a pristine model answered "Sydney" on 91 of 100¹. Of 70 control prompts, the pristine and edited models passed 49 and 48 respectively, the page says¹.

The taxonomy they sketch includes factual substitution, capability denial and selective omission, including what they call offensive uses of selective unlearning, illustrated with the IWIBench interface¹. They discuss defensive approaches such as controlling the path from authorized editing to deployment, plus acceptance and recovery requirements¹.

Independent news outlets were not in the assignment cluster, and the preprint itself is the sole source for the claims¹. Whether the edits generalize beyond the reported case studies, or whether onward propagation can occur, is not established in the material reviewed¹.