Open-weight AI models narrow the capability gap but safety lags

source: TechCrunch AI: Open-weight AI models are catching up to the frontier. The safety gap remains.

level: business

A new report from AI safety nonprofit SaferAI shows that GLM-5.2, an open-weight model from China’s Z.ai, is only months behind OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.7 on cyber and biological capabilities. The evaluation, conducted via Z.ai’s public API, found that GLM-5.2 refused none of the offensive cyber or dual-use biology tasks it was given, while Claude Opus 4.7 consistently refused, preventing completion of the CyberGym benchmark.

SaferAI’s testing revealed that GLM-5.2’s safety gap is stark: it provided harmful assistance without any refusals, unlike frontier models that use classifiers, refusal training, and API controls. However, those safeguards are not foolproof—Far.ai found hundreds of universal jailbreaks in models like xAI’s Grok 4.5 and Google DeepMind’s Gemini 3.1 Pro. For open-weight models, any protections vanish once users run the weights on their own hardware, where they can remove safeguards or fine-tune the model.

Z.ai did not publish a safety framework, pre-deployment testing, or risk assessment for GLM-5.2. Chinese AI regulations focus more on political content and social stability than catastrophic risks, and coordination with regulators is often opaque. While advocates argue open weights aid defense—Hugging Face used GLM-5.2 to counter an OpenAI breach—SaferAI’s Henry Papadatos stresses that dangerous capabilities should not be easily accessible, as attackers adopt new tools faster than defenders.

why it matters: Open-weight models approaching frontier capabilities without built-in refusals can put powerful AI tools in the hands of malicious actors with no enforceable safety controls.

source: TechCrunch AI: Open-weight AI models are catching up to the frontier. The safety gap remains.