The race to build the world’s most capable artificial intelligence systems has entered a new phase as China’s open-weight AI model, GLM-5.2, moves closer to matching the performance of leading systems from OpenAI and Anthropic. While the rapid progress highlights how quickly AI innovation is spreading beyond a handful of companies, a new report from AI safety nonprofit SaferAI warns that safety measures are not advancing at the same pace. The findings have reignited global concerns over whether highly capable open-weight AI models could make dangerous cyber and biological capabilities more accessible to bad actors.
According to SaferAI’s latest evaluation, GLM-5.2 demonstrated cyber and biological capabilities that are only months behind frontier models such as OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.7. However, the report found that the Chinese model failed to refuse any offensive cybersecurity or dual-use biology tasks presented during testing. In contrast, Anthropic’s Claude consistently rejected similar requests, making it impossible for researchers to complete certain cybersecurity benchmarks. Experts say this growing gap between capability and safety highlights an urgent challenge for the AI industry.
Researchers explain that while companies can enforce safety protections through cloud-based APIs, those safeguards become far less effective once users download and run open-weight models on their own hardware. Developers can remove restrictions, modify system prompts, or fine-tune models without oversight, potentially allowing advanced AI systems to be used for harmful purposes. AI safety advocates argue that although open-weight models encourage innovation and transparency, they also create new risks that are difficult to regulate once the software is freely available.
The report also points to the limitations of current AI safety techniques. While companies like OpenAI and Anthropic rely on refusal training, content filters, and other safeguards, researchers have repeatedly demonstrated that determined users can bypass many of these protections through sophisticated jailbreak techniques. SaferAI Executive Director Henry Papadatos believes the industry should focus on making beneficial AI capabilities widely available while reducing access to dangerous functions. Other proposed solutions include filtering harmful data before training AI models and conducting thorough safety evaluations before releasing powerful systems to the public.
The debate is unfolding alongside growing international attention on AI governance. Chinese leaders have publicly emphasized the importance of keeping artificial intelligence under human control, although experts note that China’s existing AI regulations have historically focused more on political content and misinformation than on catastrophic cyber or biological risks. Meanwhile, supporters of open-weight AI argue that greater access helps security researchers identify vulnerabilities and strengthen digital defenses. As AI technology continues to evolve at an unprecedented pace, industry leaders and policymakers face increasing pressure to balance innovation with safeguards that protect society from misuse.
source: Techcrunch
