Open-weight artificial intelligence models are rapidly catching up to industry leaders in capability, yet a significant safety gap continues to widen. A recent evaluation by the nonprofit SaferAI reveals that China’s Z.ai model, GLM-5.2, closely trails the top systems from OpenAI and Anthropic in advanced cyber and biological tasks.
Unlike closed-weight counterparts that rely on strict API controls and refusal training, open-weight models lack enforceable safeguards once downloaded onto local hardware. Researchers found that the tested model did not refuse hazardous cybersecurity or dual-use biological prompts, highlighting the inherent risks of distributing powerful AI weights without robust pre-deployment risk assessments.
As these technologies approach frontier levels, policymakers and researchers face mounting pressure to address how society can mitigate catastrophic risks. While filtering training data offers limited help for biological knowledge, securing coding capabilities without enabling cyber threats remains a complex challenge for developers worldwide.
- Open-weight AI models are rapidly closing the performance gap with industry leaders.
- The GLM-5.2 model showed zero refusals for hazardous cyber or biological prompts.
- Local execution allows users to easily strip away built-in safety mitigations.
- Experts emphasize the urgent need for better pre-deployment risk frameworks.
Sources:
