GLM-5.2: China upends the United States’ strategy to lock down dangerous AI

The United States is trying to keep its most dangerous AI systems under tight control for fear they could be used in cyberattacks. China, however, has just blown a hole in that approach - and the release of GLM-5.2 makes it plain.

Washington clamps down on Anthropic’s Fable 5 and Mythos 5

In June, the Trump administration sent shockwaves through Silicon Valley by ordering Anthropic to urgently pull its two newest artificial intelligence models, Fable 5 and Mythos 5. The move followed alarm at the highest levels of government after researchers demonstrated that the AI’s safety barriers could be bypassed.

Built to write code on the scale of an entire software product, Claude Mythos proved strikingly effective at uncovering previously unknown security weaknesses. And with officials worried that foreign powers could weaponise the technology, Washington responded forcefully by blocking exports.

Mythos 5 was eventually allowed back in early July, but under extremely strict conditions: access is limited to roughly 100 carefully vetted US organisations, including government agencies and banks. Even so, China appears to have a similarly formidable model of its own.

GLM-5.2: the Chinese equivalent with no guardrails

The start-up Z.ai - formerly known as Zhipu AI and widely seen as one of China’s “AI tigers” - has now released GLM-5.2. According to experts, this huge model delivers outstanding programming performance, putting it in direct competition with the leading US systems.

Open-weight release under the MIT licence

Unlike Anthropic, Z.ai has made GLM-5.2 available as open-weight software under the MIT licence: its code and internal parameters can be downloaded by anyone. That means the model can be run locally on private machines - a level of freedom that deeply worries Western specialists, who have little ability to contain the spread of such capability.

Experts are alarmed

Semgrep tested the model and wryly headlined its findings: “We have Mythos at home”. Their benchmarks indicate that the Chinese AI is highly effective at spotting software vulnerabilities, even outperforming certain specialised US agents - at a remarkably low cost of 17 cents per bug found. Graphistry, for its part, says this is the first freely accessible model to deliver a “frontier” level cyber experience. The threat is no longer theoretical.

Worse, the damage may already be done: hackers are reportedly already sharing methods for fully jailbreaking the model on Russian-language forums - a development that is fuelling the bleakest fears around cyber security.