Naval: Hold hosts liable for 'dangerous' or weakly-guarded open source models

naval · x · 2026-09-16

Naval extends his frontier-pacing argument: having proposed fully holding labs liable for their models' behavior, he now argues that for "dangerous" or weakly-guarded open source models, liability should extend to the hosts distributing them. End users who jailbreak models to cause harm still bear liability themselves. It's a liability-allocation framework for AI governance spanning labs, open source hosts, and end users.

Original post →

More from AGI Musings

AGI Musings channel →