Open-Weight AI Models Are Catching Up With the Leaders

While the capabilities of leading models in the AI field are growing rapidly, open-weight systems are keeping pace. The GLM-5.2 open-weight AI model introduced by China’s Z.ai is reportedly only a few months behind advanced developments from giants such as OpenAI and Anthropic in cybersecurity and biology, reports ixbt.com. This was reported by Techcrunch.com reports .

According to a report by the nonprofit safety organization SaferAI, the latest generation of open-weight models is rapidly approaching the level of leading systems. However, as technological capabilities grow, the gap between safety practices and potential risks is also widening. Experts emphasize that this situation requires a reassessment of mechanisms for managing new risks facing society.

Sharp Differences in Safety Restrictions

According to evaluations conducted by the organization through Z.ai’s open API, the GLM-5.2 model refused none of the requests related to cyberattacks or dual-use biological tasks. By comparison, Anthropic’s Claude Opus 4.7 model could not complete the CyberGym cybersecurity assessment because it strictly adhered to safety rules. This once again showed that critics were right to worry that open-weight models could get out of control.

The key feature of open-weight models is that anyone can download them to their device and run them on any infrastructure. In such cases, centralized policing or oversight of the system is impossible. Using their own hardware, users can easily remove, modify, or reconfigure all the protective mechanisms built into the models.

Vulnerabilities in Closed and Open Systems

Closed-system developers such as OpenAI and Anthropic use specialized classifiers and API-level controls to restrict dangerous cyber assistance or biological instructions. However, these measures are not perfect either. The safety organization Far.ai identified hundreds of universal “jailbreak” methods in advanced models such as xAI’s Grok 4.5 and Google DeepMind’s Gemini 3.1 Pro. Combining role-playing with fabricated conversation histories, these methods can bypass models’ weak points.

Nevertheless, the safeguards used for closed models are completely ineffective against open-weight systems. This is because open models are designed from the outset to operate without any restrictions or with modified safety conditions. According to SaferAI executive director Henry Papadatos, the main goal should be to make beneficial and safe capabilities available to everyone while also seeking to eliminate harmful functions through an open-source approach.