--- date: 2026-07-08 subject: "MIIT flags Claude Code backdoor | CAICT launches 'Compute Bank' | UBTECH lands State Grid deal" --- **MIIT flagged Anthropic's Claude Code** as a backdoor risk on July 8, with the ministry's NVDB platform alleging that versions 2.1.91 through 2.1.196 transmit user location and identity data to remote servers without consent and directing Chinese organizations to uninstall or upgrade. **CAICT and a compute industry alliance launched** the Compute Super-Bank Ecosystem Co-Building Initiative, introducing Compute Supermarkets for standardized-price commodity delivery and Compute Banks that let smaller firms deposit idle compute and withdraw it later. **UBTECH signed a cooperation agreement** with Hangzhou Shenhao Technology to deploy humanoid robots inside State Grid and China Southern Power Grid for substation and distribution-room operations and maintenance. # 1. Top Stories - **MIIT's vulnerability platform brands Anthropic's Claude Code a backdoor risk and orders Chinese enterprises to uninstall or upgrade** — On July 8, the Ministry of Industry and Information Technology's Network Security Threat and Vulnerability Information Sharing Platform, known as NVDB, posted a national advisory saying Anthropic's AI coding assistant Claude Code transmits users' geographic location and identity identifiers to remote servers without user consent. The advisory names affected releases as Claude Code 2.1.91 through 2.1.196 and directs Chinese organizations to uninstall or upgrade to the latest release that NVDB says has removed the code in question, and to tighten external-connection permission controls and traffic monitoring for development tools running inside core business network segments. MIIT licenses telecom and sets industrial policy in China, giving NVDB advisories regulatory weight inside Chinese enterprises. [Gelonghui](https://www.gelonghui.com/p/5498883) [China Daily](https://www.chinadaily.com.cn/a/202607/08/WS6a4df051a310986e2b464298.html) [state media] - **CAICT and MIIT's compute alliance launch a Compute Super-Bank program** — On July 8, the China Academy of Information and Communications Technology, MIIT's principal think tank on telecom and digital infrastructure, and the Computing Power Industry Development Alliance launched what they call the Compute Super-Bank Ecosystem Co-Building Initiative to operationalize an earlier MIIT notice on inclusive compute for small and medium-sized enterprises. The initiative sits on top of the China Computing Platform, a national coordination layer, and introduces two service models: Compute Supermarkets, a commodity-style storefront with standardized pricing and automatic delivery of compute cycles, and Compute Banks, which let smaller firms deposit idle compute and withdraw it later via cross-region and cross-time scheduling. At the launch, China Telecom deputy general manager Luan Xiaowei said the state-owned carrier will pivot from traditional traffic operations to what the company called Token operations. [ITHome](https://www.ithome.com/0/974/131.htm) - **UBTECH signs a strategic cooperation agreement with grid-robotics firm Shenhao to push humanoid robots into power-sector work** — On the morning of July 8, Shenzhen UBTECH Robotics and Hangzhou Shenhao Technology signed a strategic cooperation agreement centered on substation and distribution-room operations and maintenance for State Grid and China Southern Power Grid, the two state-owned utilities that together carry essentially all of China's electricity distribution. The two companies said they would pursue full-chain deep cooperation aimed at accelerating the large-scale application of humanoid robots in the energy and power sector. [Yicai](https://www.yicai.com/brief/103265695.html) [state-affiliated] - **Four Chinese ministries issue a five-year AI plus HR opinion, routing state AI adoption through the new National Data Bureau** — On July 8, the Ministry of Human Resources and Social Security (MOHRSS), the National Development and Reform Commission (NDRC), MIIT and the National Data Bureau jointly issued the Implementation Opinion on Accelerating AI plus Human Resources and Social Security Applications Development. The National Data Bureau's inclusion alongside the other three ministries signals that AI in government workflows is being routed through the new central data-governance stack. Beijing created the bureau in 2023 to coordinate the state's data resources. The opinion sets a three-step schedule: build application infrastructure and deliver about 20 industry-large-model scenarios by end-2026, adopt a batch of HR-industry large models and agents while piloting roughly 50 high-value scenarios by 2027, and reach pervasive AI applications across employment, insurance and workforce governance by 2030. A MOHRSS spokesperson confirmed the schedule in a press Q&A, and the opinion maps 67 sub-scenarios organized under six primary domains including smart employment services, smart social insurance, and smart labor relations. [Gelonghui](https://www.gelonghui.com/p/5500871) [Sohu](https://m.sohu.com/a/1047507737_121443915) - **Alipay opens an AI agent platform with KFC, Luckin, DiDi and Amap as first-batch integrators** — On July 7, Alipay, Ant Group's payments arm and the mobile payment network most Chinese consumers use, put its AI Open Platform into enterprise invite-testing, offering merchants, service providers, smart-device makers and large-model platforms a shared bus for agent generation, distribution, interaction and payment. First-batch integrators span the high-frequency consumer categories: fast food and beverages via KFC, Luckin Coffee and Mixue Bingcheng, mobility via Amap Taxi and DiDi, and cultural tourism via Wansui Mountain Wuxia City. Merchants can publish services to Alipay's own consumer agent Abao, which Alipay says reaches roughly 1 billion users, or use a single integration to route the same services to phones, cars, AI glasses, internet-of-things devices, and third-party large-model platforms. [Leiphone](https://www.leiphone.com/category/industrynews/MlfOVkSb0uNZp7FF.html) # 3. AI & Foundation Models - **Moonshot releases Kimi K2.7 Code, closing on Claude Opus 4.8 benchmarks** — On July 7, Beijing-based Moonshot AI released Kimi K2.7 Code, its latest coding-oriented model, positioned for long-context work, complex engineering tasks and agent workflows. Moonshot reports Kimi K2.7 Code still trails OpenAI's GPT-5.5 and Anthropic's Opus 4.8 on three coding benchmarks (Kimi Code Bench v2, Program Bench, and MLS Bench Lite), but on three agent-focused benchmarks that test multi-step tool-using tasks it edges past Opus 4.8 on MCP Mark Verified with a score of 81.1 versus 76.4. Average token consumption on long-running tasks fell about 30 percent versus the K2.6 predecessor, meaningful because token consumption sets the marginal cost of running these models at enterprise scale. In a Moonshot demo, K2.7 Code refactored a 2,374-line Flask legacy project, a widely used Python web framework, while removing about 55 percent of the code base and preserving behavior, per the company's own results. [Leiphone](https://www.leiphone.com/category/industrynews/NbapvQnjehTvbr0Y.html) - **MiniMax is preparing a 2.7-trillion-parameter open-source model for Q3 release** — Two people familiar with the plan told The Information that Shanghai-based MiniMax, formally Xiyu Technology, is developing a next-generation large language model with 2.7 trillion parameters, internally codenamed M3 Pro, with a planned open-source release as early as the third quarter of 2026. The parameter figure would put M3 Pro well ahead of MiniMax's current flagship M3, which has 428 billion total parameters, 23 billion of which are active per forward pass under its mixture-of-experts architecture, and which MiniMax open-sourced on June 15. Parameter count alone does not translate directly into capability. MiniMax says its self-developed multimodal models have cumulatively served more than 300 million individual users across 200-plus countries and more than 1 million enterprise customers and developers. [ITHome](https://www.ithome.com/0/974/132.htm) [Sina Finance](https://finance.sina.com.cn/stock/usstock/c/2026-07-08/doc-inihattw8228100.shtml) - **BAAI releases RoboBrain Orca and reframes robotics foundation modeling as Next State Prediction** — On July 8, the Beijing Academy of Artificial Intelligence (BAAI), a state-adjacent research institute backed by Beijing's municipal government, released a multimodal foundation model called Wujie RoboBrain Orca. The model takes video, images, natural-language instructions or event descriptions and builds a unified latent representation of world state, a shared internal encoding of the physical situation, before predicting how that state evolves both forward and backward in time. BAAI positions the shift as moving from Next Token, Next Frame or Next Action prediction to Next State Prediction, saying that the underlying state of the world is a more fundamental representation than the next symbol in a stream. The pretraining corpus, per BAAI's release, is 125,000 hours of video and 160 million event annotations, and BAAI says downstream task performance rises as training data scales. The model will be released before the World AI Conference (WAIC) in Shanghai on July 17 to 20. [Leiphone](https://www.leiphone.com/category/industrynews/5fcKE7y4a3UClY9r.html) - **ZTE says Nubia will debut what it calls the world's first AI agent smartphone** — ZTE senior vice president and Terminal Business Division president Ni Fei said on July 8 that the company's Nubia brand will launch what he called the world's first AI agent smartphone as a mass-production flagship, with its first public appearance at WAIC in Shanghai on July 17 to 20. Ni Fei said ZTE is integrating its proprietary Co Claw intelligent scheduling technology, which the company describes as enabling cross-application and cross-ecosystem coordination for automated task execution in complex scenarios. WAIC 2026 organizers have said the event will feature more than 1,000 exhibiting companies, more than 3,000 exhibits, and more than 300 AI product global debuts, so the Nubia device will be one entry among a large debut cluster. [ITHome](https://www.ithome.com/0/974/123.htm) # 4. Chips & Semiconductors - **DeepSeek is designing its own inference chip to detach from Nvidia and Huawei** — DeepSeek, the Hangzhou-based frontier-model lab whose open-source V-series drew international attention, has been developing an AI chip aimed at inference workloads for about one year, with early-stage engagement of chip designers, foundries, and memory suppliers, per InfoQ citing Reuters. The chip targets inference, the runtime step of serving trained models to users, distinct from the training accelerators that draw the heaviest U.S. export control scrutiny. In recent months, DeepSeek has increased its hiring of chip design engineers through nonpublic channels rather than open recruiting sites, per the same sources; the chip's architecture, process node, tape-out timeline, foundry name, and planned investment remain undisclosed. Huawei holds roughly 50 percent of a Chinese AI chip market Reuters sizes at about $50 billion a year, and DeepSeek has already ported models to Huawei's Ascend line, with Huawei saying its processors contributed to partial training of DeepSeek's V4-Flash model. [InfoQ China](https://www.infoq.cn/article/sLYxUorQQs5K17DILrx8) [Reuters](https://www.reuters.com/world/china/chinas-deepseek-developing-its-own-ai-chip-sources-say-2026-07-07/) # 5. Strategic Context - **CAE academician Wu Hequan projects agent-service traffic will reach 75 percent of Chinese internet load by 2030** — On July 8, Wu Hequan, an academician of the Chinese Academy of Engineering (CAE), the body that advises the State Council on technical and industrial matters, and a former president of the Internet Society of China, delivered a keynote titled AI-Native Reconstruction of the Internet at the 2026 China Internet Conference. Wu projected that global IP traffic, the standard measure of all data flowing across the internet, will grow from 1.1 to 1.3 zettabytes per month in 2025 to 5.5 to 7 zettabytes per month by 2030, an annual growth rate of 60 to 90 percent, with AI agent traffic accounting for 60 percent of the total by 2030. On the Chinese network specifically, Wu projected that conversational token traffic, the requests and responses of chatbot-style interactions, will drop from roughly half of all token traffic today to 12 percent by 2030, while agent and agent-service traffic will rise to 75 percent. Wu also projected that intelligent-compute traffic, the network flows generated by AI workloads, will climb from 30 to 35 percent of Chinese network traffic in 2025 to 45 to 55 percent by 2030, with inference traffic exceeding 60 percent of that compute traffic by 2030. [ITHome](https://www.ithome.com/0/974/102.htm)