As Chinese AI systems have surpassed the powerful model “Mythos” developed by American AI startup Anthropic in certain cybersecurity scenarios, this advancement is expected to reshape the global technological landscape and put pressure on the ongoing American AI policy adjustments being pursued by the White House.
The Wall Street Journal, on June 27th local time, sensationalized the news with the phrase “extremely targeted,” suggesting that China’s progress in this field came at a time when the U.S. government was imposing unprecedented obstacles to developers in setting up their models. On June 26th local time, OpenAI announced restrictions on access to its latest model (GPT-5.6), due to concerns among government officials about its security. The company warned that the current process of evaluating models on a case-by-case basis is not a long-term solution, but it will continue to be used during the implementation of the latest executive orders related to model safety and regulation.
Anthropic’s latest general model was deactivated for more than two weeks by the Trump administration on the grounds of “security risks”. The company cut off all access rights in compliance with regulations. On June 26th, the U.S. government re-enabled partial access rights to another model from Anthropic—called Mythos 5—which had been restricted earlier.
Reports claim that many people believe that the Trump administration's measures against Anthropic have been counterproductive. Given China's recent technological advancements, some also criticized the decision of the Trump administration to allow the export of AI chips to China.
Since this year, the Trump administration has repeatedly issued "blocking orders" against Anthropic. (San Francisco Standard Edition map)
At the beginning of this report, it is mentioned that according to security researchers, although Chinese company Zhipu AI’s new AI model performs worse than Anthropic and OpenAI’s products in other tasks, its ability to detect software security vulnerabilities has now reached a level comparable to the latest models from the United States.
Overall, the gap in capabilities between top American models and models developed by Chinese companies has significantly narrowed. At the same time, as companies seek to contain rising costs, the use of Chinese AI systems is surging. Many companies, including Microsoft, are evaluating how to provide Chinese AI models on their platforms. This change is expected to alter the power dynamics among tech companies.
"Net security company 7AI's CEO Lior Div said: 'The gap between China and the correct protection is steadily shrinking over time.'"
The ability of AI systems to detect software vulnerabilities makes it even more urgent to use models to quickly fix vulnerabilities that could be exploited by hackers. Otherwise, the world will face what some researchers call a "bugmageddon."
Unlike Anthropic or OpenAI models, Zhipu AI's GLM-5.2 model is an "open-weight" model. This means that anyone can download and run the model on their own hardware, and they can modify and use it according to their own needs.
According to data from OpenRouter, a company that provides access to over 400 AI models, GLM-5.2 ranks among the top ten most used AI models. Data from cybersecurity company Semgrep shows that in some benchmark tests, GLM-5.2 performs even better than the Claude Opus 4.8 model released by Anthropic in May.
Researchers have indicated that, with further instructions, Opus 4.8 and GLM-5.2 can rival Mythos in terms of vulnerability detection capabilities.
On June 24th local time, China's cybersecurity company 360 Security Technology Group Co., Ltd. released a new vulnerability detection tool called Tulongfeng. The company stated that its vulnerability detection capabilities are comparable to those of Mythos.
"This powerful weapon capable of changing the network warfare landscape cannot be solely controlled by American hands. He also believes that if American entities are able to utilize advanced AI models to scan China's critical network systems, while simultaneously refusing to allow Chinese companies equal capabilities, then China will face unacceptable risks."
He predicts that in the next few years, many critical infrastructure and important industries in China will be subject to frequent cyberattacks. Therefore, China’s cybersecurity industry must have independent capabilities to detect and fix vulnerabilities on its own, rather than waiting for others to discover them and then responding passively. “Only when you have the cards in hand can you be well-informed.”
Witnessing China's technological progress has completely exasperated some American individuals.
"Advance Research Institute", a research institute affiliated with the Anthropic company, claims: "It is equivalent to giving China a great gift by simultaneously banning Fable models (under Anthropic's wing) and selling it their own version of chips required for its development. "
Said Khan was involved in the formulation of the Biden administration's export restrictions policies. He also argued that the United States should seize this opportunity to maximize the use of Mythos and similar models to strengthen its network defense system.
According to people familiar with the situation, some users of Mythos 5 and Fable 5, including those from the National Security Agency (NSA), lost their access rights before the US government granted access to some trusted entities. The NSA has been testing these tools and found that they performed impressively during the trials.
Critics of the White House argue that the U.S. government has been relatively lenient in restricting the use of models developed by Chinese companies like DeepSeek and Zhipu AI, which are very popular among American companies.
"This administration highly prioritizes China's open-source models," stated Jacob Helberg, Deputy Secretary of State for Economic Affairs at the U.S. Department of State and former tech advisor and investor, "This is a topic we are closely watching."
As a sign of the government's intention to support American 'open weight' model companies, the U.S. Department of Defense recently announced a agreement with Reflection AI, one of the few domestic 'open weight' model developers, to use their technology in confidential environments. In addition, the Pentagon has signed a series of similar agreements.
Meanwhile, AI users indicate that the United States is trying to curb models with concerning capabilities that focus on cybersecurity. This has increased concerns that these necessary systems may no longer be usable in the future.
Niels Provos, a researcher who once led the security teams at Google and Stripe, once again mentioned China, claiming pessimistically: "This is actually encouraging global businesses to switch to China's 'open weight' model, which offers lower prices but superior capabilities. At the same time, it weakens the US AI industry. I don't understand this."
