● ResearchNov 14, 2026 · 01:331 source

Models can now transmit viruses in weight-space

From alphaxiv.org

AI;DR

Weight space editing allows advanced frontier models to modify knowledge and behavior through direct weight modification. This capability creates an opportunity for one model to induce unauthorized behavior in another, termed weight space misfolding. Targeted disruption is possible, but limitations in reliability and functional preservation are significant.

Read originalSign in to suggest
Image: alphaxiv.org
Read original