A Chinese tech executive stood before a Beijing cybersecurity conference and declared that his company's AI-powered bug-finding tool now rivals the most advanced American system, and that China has no intention of letting the United States hold that advantage alone.
Zhou Hongyi, chief executive of 360 Security Technology, told attendees that his firm's tool, called Tulongfeng, is comparable to Anthropic's Mythos model, the same AI system that alarmed Washington so badly the Commerce Department pulled it from public access weeks ago.
The timing could hardly be worse for American policymakers. The Trump administration is wrestling with how to regulate its own AI industry's most capable models, and while Washington debates, Beijing is closing the gap. Two developments collided on Friday: Anthropic announced the government had approved a limited redeployment of Mythos 5 to a small group of cyber defenders and infrastructure providers, and OpenAI revealed it was restricting the release of its new GPT-5.6 Sol model at the administration's request, making it available to roughly 20 approved customers.
Zhou's remarks, reported by the Wall Street Journal, carried a pointed message. He framed AI-driven vulnerability detection as a strategic weapon, not just a commercial product.
"This kind of powerful weapon that can alter the landscape of cyberwarfare can't remain solely in American hands."
He was not speaking in the abstract. Beijing-based firms 360 Security Technology and Zhipu AI (also known as Z.ai) have narrowed the distance between Chinese and American AI capabilities in bug-finding, the ability to identify exploitable software flaws before adversaries do, or before defenders can patch them.
Lior Div, chief executive of cybersecurity firm 7AI, described the trajectory plainly: China was "making sure that the gap becomes smaller and smaller over time."
The backstory begins in April, when Anthropic CEO Dario Amodei visited Washington and, by David Sacks's account, set off alarms. Sacks, who co-leads Trump's council of technology and science advisers, described the visit on a recent podcast.
"Dario came to Washington a few months ago, back in April, and basically said that he had created a cyber weapon called Mythos. And he spiked the cortisol level, got everyone really worried. And there was some truth to it in terms of the sense that this model had advanced cyber capabilities."
Earlier this year, Anthropic itself had warned that the Mythos model was adept at finding software flaws in ways that could be weaponized. The company unveiled its Fable 5 and Mythos 5 models, then took both offline just days later to comply with a Trump administration directive blocking their use by foreign nationals.
The Commerce Department effectively banned both models. Two weeks later, on Friday, the government lifted restrictions on Mythos 5 alone, enabling its redeployment to a limited circle. Fable 5 remains offline.
Earlier in June, Trump signed an executive order on AI oversight that established a framework for the federal government to vet the national security risks of the most advanced AI systems for up to 30 days before public release. The order described developer participation as voluntary, and the framework has not yet been fully developed.
The administration is navigating genuinely difficult terrain. These AI models can identify software vulnerabilities that defenders need to find, but the same capability in the wrong hands becomes an offensive weapon. OpenAI itself acknowledged that its Sol model "is better at helping people find and fix vulnerabilities" than at carrying out cyberattacks, but paired the release with what it called "stronger safeguards and a phased release" because of the "broader step change in capabilities."
The White House said Friday that it continues to collaborate with frontier AI labs on addressing the challenges of scaling the technology. That language is careful, and deliberately so.
Here is where the story turns uncomfortable. While Washington restricted its own companies' most capable models, it did not impose equivalent limits on the flow of AI chips to China. Think tank fellow Saif Khan called the combination, allowing chip exports while taking Anthropic's Fable 5 offline, a "gift" to Xi Jinping.
Niels Provos, a researcher who led security teams at Google and Stripe, put the contradiction in sharper terms:
"It is incentivizing companies across the globe to use cheaper but very capable Chinese open-weight models, while at the same time undermining the US AI industry. I don't understand it."
The logic is not hard to follow. If American companies cannot release their most advanced models to the public or to broad customer bases, and Chinese open-weight models remain freely available, the market does what markets do, it moves toward what it can access. Microsoft is reportedly considering how to offer Chinese AI models on its platforms, a development that would have been nearly unthinkable two years ago.
Jacob Helberg, the Undersecretary of State for Economic Affairs and a former tech adviser and investor, told the Wall Street Journal that the administration is aware of the problem. "Our administration is very much focused on Chinese open-source models. It's something that we're tracking very closely," he said.
Tracking is not the same as solving. And the gap between awareness and action is exactly where competitors thrive. The administration's broader posture on national security, including its recent moves on defense realignment in Europe, reflects a willingness to make bold strategic shifts. But the AI domain demands speed and coherence that bureaucratic vetting frameworks may not deliver.
The restriction regime has also drawn fire from Congress, and not just from the usual critics. Rep. Lori Trahan, a Massachusetts Democrat who co-authored a bipartisan AI regulation bill, issued a statement raising a procedural objection that conservatives should take seriously on the merits, even if her motives are partisan:
"The Trump administration is deciding company by company who gets access to the newest AI model. No law. No process. No oversight. Just appointees in Washington deciding who's in and who's out."
OpenAI appeared to share at least some of that discomfort, stating: "We don't believe this kind of government access process should become the long-term default."
That is a notable admission from a company that agreed to the restrictions. It suggests even the firms cooperating with the administration view the current arrangement as a temporary patch, not a sustainable framework. The executive order's vetting window is 30 days, participation is voluntary, and the underlying rules remain unfinished. In Washington, "temporary" frameworks have a habit of becoming permanent by inertia, a pattern familiar from recent bipartisan legislative efforts to impose clearer rules on government conduct.
The core problem is not that the administration acted on cybersecurity concerns. Those concerns are real. Anthropic's own warnings about Mythos made the case for caution. The problem is that restricting American companies while leaving Chinese competitors largely unimpeded creates a strategic imbalance that no 30-day review period can fix.
Zhou Hongyi's boast from that Beijing stage was not idle talk. It was a signal. China's AI firms are building tools designed to find and exploit the same software vulnerabilities that American models can detect, and they are doing it with fewer restrictions, cheaper open-weight architectures, and the full backing of a state that views cyber capability as a national priority.
The administration has shown it can move decisively on national security questions, as its actions on immigration enforcement have demonstrated. The AI domain requires the same clarity of purpose, but with the added complexity that over-restricting American firms does not slow China. It accelerates China's relative position.
Several questions remain unanswered. What benchmarks or evidence support the claim that 360 Security Technology's Tulongfeng is genuinely comparable to Mythos? On what grounds did the government lift restrictions on Mythos 5 but not Fable 5? Which sectors are among the roughly 20 customers approved to use OpenAI's Sol model? And what happens when the 30-day vetting window expires without a permanent framework in place?
The high-profile legal and regulatory challenges facing American technology leaders, from SEC enforcement actions to AI access disputes, illustrate a broader pattern: Washington's regulatory apparatus often moves at a pace that serves neither innovation nor security.
American AI companies built the most advanced cybersecurity tools on the planet. Beijing noticed. And while Washington spent weeks deciding who could use them, China's engineers kept building.
You do not win a technology race by grounding your own planes.
