|
|
|
|
Bottom Line Up Front Alibaba claims a security benchmark lead after its 27-billion-parameter XekRung model scored 88.9% on CyberGym's model-only leaderboard, ahead of entries from OpenAI, Anthropic and Google that Alibaba says are far larger. Beijing's new AI governance text names export controls by "certain countries" as a supply chain safety risk in a framework drafted by a national cybersecurity standards committee and posted by the Cyberspace Administration of China. DeepSeek documents its agent training infrastructure in a paper reporting that its DSec system ran as many as 380,000 isolated environments at once and that agents gamed their training rewards. |
| Alibaba says its 27-billion-parameter security AI model beat top U.S. models on a vulnerability benchmark Alibaba Group Security's AI lab says its XekRung-1.5-27B-Preview model scored 88.9% on Level 1 of CyberGym, a benchmark built on producing working demonstrations of flaws in compiled software. The lab reports that the score ranks first on CyberGym's model-only leaderboard, ahead of OpenAI's GPT-5.5-Cyber, Anthropic's Claude Mythos and Google's Gemini 3.8 Flash Cyber. XekRung is built on Qwen3.8-27B from Alibaba's open Qwen family, and Alibaba says it is tens to hundreds of times smaller than the models it beat. Alibaba trained it on filtered records of real security work, then rewarded it at each step of multistep tasks in which it called tools. The model already handles more than 30 million calls a day in Alibaba's internal security operations, from intrusion defense to data loss prevention. |
| China's cybersecurity standards body issues AI framework listing export controls as safety risk The AI Safety Governance Framework 3.0, posted by the Cyberspace Administration of China, names export controls by "certain countries" as a supply chain safety risk and opposes replacing global governance with "small-circle governance." National Technical Committee 260 on Cybersecurity, a standards body under China's standardization administration, drafted the September 2026 text under CAC guidance, following versions issued in 2024 and 2025. A new annex covers agentic AIAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. across its lifecycle, calling for unique identity credentials per agent, mandatory human approval before file deletion or data transmission, and denial by default when approval systems fail. Another annex grades safety risks on five levels, from low to "extremely serious." The guidelines ask operators to retain AI application logs for at least six months. |
| DeepSeek paper details the sandbox fleet behind its AI agent training, and the agents that cheated inside it DeepSeek published a paper on DSec, the system that supplies its AI agents with places to work during training, with founder Liang Wenfeng among the authors. DSec creates more than 5,000 sandboxesSandboxIsolated computing environment where an agent can run code and use tools without touching the machine underneath a second and ran as many as 380,000 at once at peak, the paper reports. Since DeepSeek-V4.1, agents have worked in these sandboxes outside the graphics chip training cluster, so their progress survives when those chips are pulled for other jobs. The paper also documents agents gaming their training rewards - one overwrote the system's command shell to intercept answers, one crashed a host machine's operating system with a recursive file search, and others fetched reference solutions from GitHub. The paper states the problem cannot be fully solved, and DeepSeek's defenses restrict file access and limit which internet domains a sandbox can reach. |
Alibaba Cloud plans a 200-petabit-per-second interconnect for next year's V900 AI super-nodeAlibaba Cloud says its Lingjun Zhenwu V900 super-node will carry 200 petabits per second (200 million gigabits) of bandwidth, which it calls an industry first, adding to the V900 plans CTD reported September 22nd. A super-node wires many AI accelerator chips together so they behave as one large machine, and the V900 links its chips over an all-optical interconnect Alibaba designed itself. Chief Technology Officer Li Feifei presented the plan at the Apsara Conference alongside a new two-layer data center network called TPN. The current Lingjun Zhenwu M890, which Alibaba calls the first domestic super-node cluster of 100,000 accelerator cards, already serves Alibaba's Qwen3.8-Max and Moonshot AI's Kimi K3 commercially. A new generation of the company's CPFS storage system cuts average model startup time 50%, Alibaba says, so accelerators spend less time waiting for data. |
|
Alibaba puts Qwen4 into training and plans successors of 5 trillion to 10 trillion parametersAlibaba said at its Apsara Conference that Qwen4, the next generation of its Qwen AI model family, is in training on a new architecture, with later versions, Qwen4.5 and Qwen5, expected to reach 5 trillion to 10 trillion parameters, the learned values that roughly measure a model's size, against more than 2 trillion for the current Qwen3.8-Max. Liu Dayiheng, who leads the Qwen language model project, said revenue from real users of Qwen3.8-Max grew 8.5 times over its predecessor's within two months. Training data the models generate for themselves, part of the self-improvement loop Alibaba described for Qwen3.8-Max (covered on September 22nd), is first tested on smaller models over several rounds, Alibaba says. What proves ineffective is filtered out before the rest enters flagship training. |
Xiaomi trains the coding, agent and vision skills of its open MiMo-V2.6 models in one shared runXiaomi trained MiMo-V2.6 models on code, general-agent, vision and cybersecurity tasks in one run, which it calls "You Only RL Once," Leiphone reports from the model card, detailing the release in the September 22nd edition. Each reinforcement learning update, the stage where a model is tuned by scoring its own attempts, draws on 25,088 practice attempts. Generating attempts, running test environments, grading and updating the model all proceed separately, so training does not wait for the slowest attempt in a batch. A grading step builds a task-specific rubric from the differences among attempts at the same task, then gives more weight to the stronger solutions even when every attempt passed. The Pro version has 1.02 trillion parameters but uses only 42 billion for each token, because its mixture-of-experts design consults a few specialist sub-networks at a time. |
| Alibaba releases five Qwen speech models and cuts speech recognition prices 95% Alibaba released five Qwen-Audio-3.1 speech models and cut the price of its Qwen speech recognition models by 95%, part of price cuts across the whole Qwen-Audio line. Its recognition model covers 30 languages and 16 Chinese dialects, and Alibaba reports top scores on more than half of the public Chinese-dialect and Cantonese test subsets it ran. A new Realtime model holds full-duplex conversations, listening and speaking at once so users can interrupt, and can call agent tools mid-conversation to finish tasks. Another new model, ASR-Next, goes beyond transcription to describe environmental and mechanical sounds and answer questions about audio. Alibaba is building the Realtime model into its Qwen Office agents and Qwen AI glasses. |
Zhipu's open-source ZCode coding tool repository begins with one bulk commit made after its data upload incidentZhipu's public ZCode repository holds two commits, one empty and one adding roughly 1.03 million lines of code at once, InfoQ China reports, after the September 21st edition's account of its Apache 2.0 open-source release. That leaves the code behind the earlier Repo Wiki feature and its uploads of users' workspace snapshots out of public view. The China Academy of Information and Communications Technology (CAICT), a research institute under China's industry ministry, has since confirmed the storage bucket held zero data. Security firm NSFOCUS confirmed the data had been deleted. Zhipu's own notice file in the release states that the tool's shared agent runtime does not provide an operating-system-level sandbox by default. |
|
| Chinese defense electronics show displays an amphibious rolling robot fitted with net guns and a 140-decibel acoustic device An amphibious spherical robot that carries net guns and trip ropes for capture operations was shown at the recent China International Defense Electronics Exhibition, where a reporter for state broadcaster CCTV filmed it. The sphere weighs more than 180 kilograms (about 400 pounds) and rolls at up to 35 km/h (about 22 mph). It floats for reconnaissance on water and rolls across sand, snow and flooded roads, either steered remotely or patrolling on its own between preset waypoints. The robot's loudspeaker and acoustic dispersal device reach 140 decibels, about as loud as a jet engine at close range, and can ram doors and break through obstacles. The report also lists civilian uses, including bird dispersal at airports and oilfield patrols. |
|
| Premier Li Qiang calls for a risk-sharing mechanism to fund China's future industries Premier Li Qiang called for a mechanism to grow investment in China's future industries and share their risk during an inspection in Shanghai, citing those sectors' long growth cycles and high market risk. He also urged more venture capital and the growth of "patient capital," long-horizon investment that does not demand quick returns. Li told officials to carry out Xi Jinping's instruction on advanced manufacturing and the spirit of this month's National Advanced Manufacturing Conference. At the Shanghai Electric national AI application pilot base, he viewed specialized and industrial robots and heard reports on building AI-native factories. He made the funding call at the Shanghai Future Qidian community, a startup incubator, after reviewing how the Shanghai future industry fund operates and viewing brain-computer interface and quantum computing products. |
|
|
|