Related: White House Lets Anthropic Restore Its Top AI Model for US Users
A Chinese artificial intelligence model is now matching one of the most tightly guarded AI systems in the United States at a task governments treat as a matter of national security, finding security flaws in software. According to a Wall Street Journal report, security researchers have found that GLM 5.2, an open model from the Chinese firm Zhipu AI, performs on par with the latest American models at detecting software bugs, and in some tests it edges them out. The finding cuts to the heart of the strategy Washington has been building to keep the most powerful cyber capable AI out of rival hands.
The comparison that matters is with Mythos, Anthropic‘s frontier cyber model, which the US government considers so capable at finding and exploiting software vulnerabilities that it has been kept under strict control. On the Semgrep benchmark, a standard test of bug detection, GLM 5.2 sometimes outperformed Anthropic’s Claude Opus 4.8, and with additional prompting both models reached a level of vulnerability detection comparable to Mythos. In one specific measure, the detection of a common class of access control flaw, independent testing put GLM 5.2’s score ahead of Anthropic’s coding tool. The Chinese model is not just close, it is competitive at the exact capability the United States has tried hardest to contain.
What makes the result more than a benchmark curiosity is how GLM 5.2 is distributed. Zhipu released it on June 13 as an open weight model, one day after the US government blocked global access to Anthropic’s Fable 5, which means anyone can download it, run it on ordinary hardware, and modify it freely. Mythos, by contrast, reaches only a vetted list of organizations and sits behind export controls that bar foreign access. A capability the United States is rationing through policy is now available to anyone in the world with an internet connection, including the adversarial actors those controls were meant to keep out.
Zhipu is not alone in pressing the point. On June 24, the Chinese firm 360 Security Technology unveiled a bug detection tool it calls Tulongfeng, which it says matches Mythos as well. Together the two announcements amount to a coordinated message from China’s AI sector that the gap with American frontier models in cyber capability has narrowed to the point of disappearing, and that the country can deliver these tools through open release rather than the gated, controlled approach the United States has chosen.
The timing turns the episode into a direct challenge to US policy. Over the past month Washington has leaned hard on export controls and access restrictions, pulling Anthropic’s most powerful models offline before allowing a limited, vetted return, and pressing OpenAI to hand out its newest system only to government approved customers. The logic behind all of it is that restricting access buys safety and strategic advantage. An open Chinese model that matches the restricted American one undermines both halves of that argument at once, since the capability is no longer scarce and the restrictions mainly handicap American companies rather than denying the tool to rivals.
There are reasons for caution before declaring the race over. Benchmark scores capture narrow slices of capability and do not always translate to real world performance, the testing here focuses on specific bug detection tasks rather than the full sweep of what a frontier cyber model can do, and a single class of result is not the same as parity across the board. Chinese labs also have an incentive to frame their models as equal to the best American systems, so the claims deserve scrutiny rather than reflexive acceptance.
Even with those caveats, the development lands at an awkward moment for the United States. The country has spent the year treating advanced AI as a strategic asset to be controlled, and the appearance of an open Chinese model that matches its most restricted system suggests that control may be harder to maintain than policymakers hoped. The question now facing Washington is whether export controls and access limits still make sense when the capability they target can be downloaded for free from a rival, or whether the contest has already moved to ground where restriction no longer holds the advantage.
Related on Entrelligence: the US lifting export controls on Claude Fable 5, Claude running on NVIDIA GPUs in Azure, and Google’s Gemma 4 passing 200 million downloads.
Your First 10 AI Skills
10 practical AI skills, copy-paste prompts and a 7-day plan to start using AI with confidence.
Download the guide →