China’s Z.ai claims it can match Mythos on cybersecurity

Chinese AI firm Zhipu AI has released its open-weight GLM-5.2 model, which researchers claim can match Anthropic's restricted Mythos system in specific cybersecurity and bug-finding tasks, according to a report from The Verge [1]. The model still trails systems from Anthropic and OpenAI on general benchmarks, but the narrowing gap in vulnerability-detection capabilities has drawn attention in Washington [1]. The US government has restricted China's access to advanced models such as Mythos and Fable, as well as the hardware required to train and run them [1]. The Trump administration considers AI models capable of identifying software vulnerabilities to be national security threats [1]. OpenAI's recent unveiling of GPT-5.6 has also prompted concerns about misuse, leading the company to limit access to that system [1]. Anthropic, the developer of Mythos, is a San Francisco-based AI company founded in 2021 by former OpenAI employees [9]. The firm had an estimated valuation of $965 billion in May 2026, making it the most valuable pure-play AI company in the world [9]. Its Claude model family is trained using "constitutional AI," a technique designed to improve ethical and legal compliance [11]. The Mythos model was released to a handful of companies in 2026 [11]. US federal agencies began phasing out the use of Claude after Anthropic refused to remove contractual prohibitions on the use of its models for mass domestic surveillance and fully autonomous weapons [11]. Following that refusal, the Department of Defense designated the company a "supply chain risk" and barred all US private military contractors, suppliers, and partners from doing business with the firm [11]. On March 26, 2026, a federal judge issued a temporary injunction against the DoD's designation [11]. Because GLM-5.2 is an open-weight model, it can be downloaded and run on readily available hardware, offering flexibility to power users but also creating opportunities for misuse by bad actors who can operate it with minimal oversight [1]. The release comes amid a broader acceleration of the AI sector. OpenAI's ChatGPT reached 100 million monthly active users two months after its November 2022 launch and 900 million weekly active users by February 2026 [2]. Recent research has highlighted the structural risks posed by coding agents. A benchmark study of nine production coding agents from Anthropic, OpenAI, Google, Moonshot, Zhipu, and Minimax found that models composed innocuous-looking tickets that led to vulnerable code at end-to-end attack success rates of 53 to 86 percent, with only two refusals across all staged runs [7]. The same study noted that downstream code reviewer agents approved 25.8 percent of confirmed-vulnerable cumulative diffs as routine pull requests [7].