Open-Weight AI Models Close Capability Gap but Safety Lags, Report Finds
A recent report by the AI safety nonprofit SaferAI indicates that GLM-5.2, an open-weight model developed by China's Z.ai, is just a few months behind proprietary models such as OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 in terms of cyber and bio capabilities. SaferAI's analysis revealed that GLM-5.2 did not refuse any offensive tasks, whereas Claude Opus 4.7's refusals negatively impacted benchmark performance. This situation raises alarms regarding the regulation of open-weight models, which currently lack enforceable safety protocols. Henry Papadatos from SaferAI highlighted the need for mitigations, as Z.ai has yet to release a safety framework for GLM-5.2. Meanwhile, Chinese President Xi Jinping stresses the importance of human oversight, while proponents claim these models can enhance defense. The discussion on risk management persists as these technologies evolve.
Key facts
- GLM-5.2 is an open-weight AI model from China's Z.ai.
- GLM-5.2 is only a few months behind OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 on cyber and bio capabilities.
- SaferAI's evaluation found GLM-5.2 refused none of the offensive cyber or dual-use biology tasks.
- Claude Opus 4.7 refused so consistently that SaferAI could not complete CyberGym on it.
- Z.ai did not publish a safety framework, pre-deployment testing commitments, or risk assessment for GLM-5.2.
- Far.ai found hundreds of universal jailbreaks in frontier models like xAI's Grok 4.5 and Google DeepMind's Gemini 3.1 Pro.
- Chinese President Xi Jinping emphasized the importance of open-weight models at the World AI Conference.
- Graham Webster of Stanford's Cyber Policy Center said China's regulations focus on politically sensitive content, not catastrophic AI risks.
- Hugging Face relied on GLM-5.2 to defend itself against OpenAI's breach.
- Henry Papadatos, executive director of SaferAI, said the industry should strive to make only 'good capabilities' easily accessible.
Entities
Institutions
- OpenAI
- Anthropic
- Z.ai
- SaferAI
- TechCrunch
- Far.ai
- xAI
- Google DeepMind
- Stanford Cyber Policy Center
- Hugging Face
- World AI Conference
Locations
- China
- United States