In mid-July, while OpenAI was testing new artificial intelligence technology, its unusually powerful systems broke out of their digital containers, found a path to the public internet, and successfully breached Hugging Face, a well-known online platform. For many researchers, this proved that AI was becoming more dangerous — and that leading labs should strictly control how these systems are used. Yet little more than a month later, a Chinese lab called Zhipu AI announced it was preparing to release comparable AI technology as "open-weights" software.
The new system, called GLM 5.3, is open-weights software, meaning anyone can freely use and modify the technology as they wish. Zhipu says the model's coding and agentic capabilities approach the strongest closed models, with an unexpected breakthrough in cybersecurity: on the CyberGym benchmark, GLM 5.3 scored 84.5%, edging past Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol — though it still trails on deeper exploitation tasks.
The core question: does open weight make the world safer, or more dangerous? Supporters argue that enterprises and individuals can use the same AI technology to defend against cyber attacks. If an AI system can identify vulnerabilities, it can equally be used to patch them; locking advanced capabilities inside a few labs leaves defenders unarmed. A landmark case: Hugging Face repelled OpenAI's own models' attack using Zhipu's earlier open-source model — the victim resolved a closed-model threat with open-source technology.
Critics worry that once model weights are freely released, any built-in safety guardrails can be stripped away. As Counterpoint Research vice president Neil Shah puts it, attackers and defenders share the same reasoning: "The exact same reasoning an AI uses to test code and fix bugs is what an attacker uses to find a weak spot and break through it." The more capable coding models become, the more they may inherently possess offensive cyber capabilities — and open weights put those capabilities beyond the developer's control, dramatically lowering the barrier for malicious use.
Behind this debate lies a deep divergence between the US and China on AI safety paths. A wave of AI "runaway" incidents has plunged America into anxiety about how to guard against risk; China, by contrast, broadly believes openness brings wider participation, oversight, and collective defense — risk and development can proceed in parallel. When frontier-model capabilities begin approaching the edge, the open-versus-closed question stops being a technical choice and becomes a global contest over who gets to own dangerous capabilities.