Policy & Safety

Open-weights AI survives its biggest security test — and its critics regroup

A year of feared misuse scenarios largely failed to materialise, shifting the policy debate from blocking releases to hardening society's defences.

By Sofia Lindgren, AI Policy Correspondent — GENEVA

GENEVA — A year after open-weights models reached near-frontier capability, the catastrophes predicted at their release have largely failed to arrive, according to a multi-government threat review published this week — a finding now reshaping one of AI policy's fiercest arguments.

The review documents real but bounded misuse — scaled phishing, disinformation content, some malware assistance — while finding that defensive applications of the same open models, from intrusion detection to scam-call interception, diffused faster than offensive ones.

Restriction advocates counter that the finding is generational: today's open models cannot yet design novel pathogens or autonomously penetrate hardened networks, and the review itself recommends capability-triggered controls rather than permanent comfort.

Enable JavaScript to read the full story on Neural Daily News.