A Chinese AI lab claims it can compete with Anthropic’s extremely powerful Claude Mythos in identifying security vulnerabilities.

      Security researchers claim that Z.ai's latest model can compete with Anthropic's Mythos in a significant area.

      For several weeks, Anthropic’s Mythos has been regarded as the benchmark for AI-driven cybersecurity. However, that lead may be diminishing. A report from The Wall Street Journal indicates that security researchers believe Z.ai’s GLM-5.2 now matches Mythos in detecting software security vulnerabilities, although it still falls behind Anthropic and OpenAI in broader reasoning abilities.

      GLM-5.2 is reducing the disparity in a crucial domain.

      The report states that researchers found GLM-5.2 performs similarly to Mythos in locating software bugs, a skill that is increasingly vital as companies scramble to address vulnerabilities before they can be exploited by hackers. The model’s open-source nature allows anyone to download, modify, and utilize it on their own systems without depender on a cloud service. This adaptability makes it appealing for businesses, but it also raises concerns that cybercriminals might exploit it for malicious activities.

      The report emphasizes that this does not imply that China has surpassed the U.S. in AI overall. GLM-5.2 still lags behind Anthropic and OpenAI in various general tasks. However, in the realm of cybersecurity, where even slight advancements can have significant real-world implications, the performance gap has notably decreased. According to benchmark data referenced by the Journal, GLM-5.2 has even outshone Claude Opus 4.8 in certain security tests, and researchers note that additional prompts enable it to achieve Mythos-level bug-detection capabilities.

      The broader narrative isn’t about who prevails; it’s how swiftly the gap is closing.

      Interestingly, this situation arises at a rather inconvenient time for the U.S. AI sector. While companies like Anthropic and OpenAI have been limiting access to their most advanced models due to national security concerns, Chinese laboratories have been moving in the opposite direction, launching increasingly sophisticated open-weight alternatives that anyone can download and utilize.

      This public discourse was apparent just days ago when Elon Musk speculated that Chinese AI labs are likely to catch up to Anthropic’s premier Fable 5 by the first quarter of 2027, particularly regarding benchmark performance. Zhipu AI founder Tang Jie swiftly countered, stating it “won’t take that long.” Musk later clarified his stance, asserting that while China might match Anthropic in benchmarks by then, achieving a comparable level of “true usefulness” would be a greater challenge, crediting Anthropic’s focus on practical intelligence.

      While acknowledging benchmark results is one thing, reaching true usefulness even by Q1 would be a significant accomplishment. Anthropic has wisely concentrated on enhancing useful intelligence, which may not be reflected in benchmarks but certainly impacts revenue.— Elon Musk (@elonmusk) June 18, 2026

      Now, The Wall Street Journal’s recent report lends more credibility to Tang’s optimism. Rather than focusing solely on coding benchmarks, it indicates that GLM-5.2 is already on par with Anthropic’s Mythos in detecting security vulnerabilities, which is arguably one of the most crucial real-world AI applications today. This does not instantaneously position China as the frontrunner in frontier AI, but one aspect is becoming increasingly difficult to overlook: the AI race no longer offers a comfortable lead for the United States.