Alibaba details new AI safeguards as Chinese developers step up test disclosures
By Eduardo Baptista and Harishankar Manoj
Nov 15 (Reuters) - Alibaba 9988.HK has disclosed new safety measures for its flagship AI model, including restrictions on retaining cyberattack-related training data, as Chinese AI developers detail more safeguards against the misuse of increasingly powerful AI systems.
The Hangzhou-based e-commerce and technology conglomerate said Qwen4-Max, unveiled on Monday, is trained to retain cybersecurity knowledge useful for detecting and fixing software vulnerabilities while excluding material intended to improve offensive hacking capabilities, Alibaba said in a statement.
The United States and China are both racing to develop increasingly powerful AI systems, but now wrestle with growing concerns that the technology could be exploited to launch cyberattacks, develop dangerous weapons, or act beyond human control — a scenario sometimes described as "rogue AI".
Alibaba's Chinese rivals Tencent and Moonshot AI have already announced safeguards, suggesting a shift towards greater transparency after an October report by research firm SemiAnalysis found that China's leading AI developers had published safety-test results for just 3.6% of 857 model releases.
Tencent said this month it had restricted cybersecurity training for its new Hy4 model to defensive uses, while Moonshot said in October it had deliberately reduced the hacking capabilities of its Kimi K3.1 Pro model.
Alibaba said Qwen4-Max had improved its score on a benchmark measuring secure code generation across functions, files and software repositories to 14.7% from 8% for its predecessor.
Alibaba said the model also had stronger safeguards against chemical, biological, radiological and nuclear threats, adding internal evaluations had found no evidence that the model could significantly enhance users' ability to carry out dangerous tasks.
The findings have not been independently verified by Reuters.
Alibaba said Qwen4-Max, built with 2.4 trillion parameters, had also reduced average task costs by around 60% in internal testing, compared with its predecessor.
The company plans to release smaller open-weight models from the Qwen4 family, allowing developers worldwide to download and modify them.