New Auto Safety Rules Draw Red Lines, Claude Bans 11.4M Accounts in Six Months: Tech Roundup

China's Ministry of Industry and Information Technology and three other agencies just dropped a sweeping draft regulation on automotive product design — sending shockwaves through the industry. Meanwhile, across the Pacific, Anthropic's data showing 11.4 million Claude account bans in six months has lit up AI developer communities. Today's tech news spans smart vehicle safety, AI governance, the personal AI agent race, and chip supply-demand dynamics — several storylines that together reveal the real anxieties running through the industry right now.

The Auto Industry Collectively Says No to "Rush-to-Market" Cars

On October 10, the Ministry of Industry and Information Technology (MIIT) and three other departments released a draft notice titled Notice on Further Strengthening Management of Automotive Product Innovation Design and R&D Testing Verification. It sets out specific testing and verification requirements covering virtual controls, door handles, new human-machine interaction systems, and large-angle reclining seats. The draft lays down several explicit prohibitions: fully hidden door handles are banned outright; safety-critical functions such as lighting, wipers, and defrosting/defogging must retain physical buttons if they use touchscreen virtual controls; production vehicles may not be equipped with foldable or flexible in-car displays; and zero-gravity seats, rotating seats, or fully reclined front seats that convert into beds may not be activated while the vehicle is in normal motion.

The timeline is clear: starting January 1, 2027, newly submitted vehicle models must include dedicated technical parameters and testing verification materials; vehicles already on sale must complete supplementary filings by July 1, 2027, and those that miss the deadline while posing safety risks must halt production and issue recalls.

Great Wall Motor's WEY brand CEO Zhao Yongpo immediately posted: "The greatest luxury isn't the sofa — it's the brakes." That line perfectly captures the regulation's core logic: at a moment when the industry has been stacking features and selling cockpit experiences, regulators are pulling the wheel back toward the safety baseline. FAW Group also stated it would "insist on not putting anything on the road before it's ready, and not launching anything that hasn't been fully validated."

Another thread worth watching: on October 9, the China Automotive Technology and Research Center (CATARC) urgently sent a survey questionnaire to major automakers inquiring about brake pedal assembly materials, with a focus on non-metallic materials — Changan Automobile had already received it by that day. Surveys like this are typically read as preparatory moves ahead of regulatory revisions, signaling that safety standards around braking will continue to tighten.

Personal AI Agents Become the New Battleground for Chinese Tech Giants

Following Meta's launch of its personal AI agent Muse, China's tech giants have moved to catch up with remarkable speed. ByteDance was pushing hard on a personal AI agent product right before the Golden Week holiday — the project is codenamed "Spell," launched in April this year, and is a continuation of the Doubao mobile assistant's personal agent direction for 2025. The product is led by the Doubao mobile assistant team, jointly developed with the Doubao conversational AI team, with the goal of launching publicly as soon as possible.

Tencent launched its cloud personal assistant LightVela in late September. Accessible via WeChat, it handles daily requests around the clock and is currently in a developer experience phase. Ant Group's "Lingguan" has also opened a small-scale beta for a new version. People close to the project revealed that the core of this upgrade is a shift from "an AI assistant that can chat and generate apps" to "an AI-native personal agent with long-term memory that can proactively advance tasks" — capable of actions like sending DingTalk messages and controlling smart home devices by calling Skills and MCP.

Lingguan 2.0's beta use cases center on long-horizon, multi-variable scenarios that are hard for ordinary users to manually track — like helping office workers monitor market movements, helping parents manage DingTalk attendance check-ins, or rerouting a trip when a flight is delayed. Looking at how these companies are defining their products, the real competition in the personal agent race is no longer about conversational capability, but about "long-term memory + proactive execution" — marking a paradigm shift from AI assistants that passively respond to agents that actively act on your behalf.

What is MCP? MCP (Model Context Protocol) is an open standard protocol proposed by Anthropic in late 2024, designed to let AI models connect to external tools, databases, and services in a standardized way. Think of it as a universal adapter between an AI assistant and the outside world — developers write the interface once according to the protocol spec, and the AI can then call DingTalk, smart home devices, calendars, and other applications without needing a custom integration layer for each one. When Ant's Lingguan 2.0 mentions "calling Skills and MCP," it's leveraging this protocol to execute proactive, cross-application tasks. Skills is an earlier concept referring to individual capability modules built into the AI assistant — things like sending a message or checking the weather. Combined, the two enable personal agents to evolve from "answering questions" to "completing multi-step, cross-platform tasks on the user's behalf" — and that's the core technical infrastructure driving competition among major players today.

Claude Bans 11.4M Accounts: AI Governance Dilemma Surfaces

A report Anthropic published in July was recently dug back up, and the numbers are striking: in the first half of 2026, Claude cumulatively banned 11.4 million accounts. Of the 398,000 appeals submitted, only 42,000 succeeded — an overturn rate of roughly 10.5%. The figures sparked million-view discussions across AI developer communities.

Anthropic didn't disclose the full breakdown of ban reasons, but listed several triggering categories: instructions involving illegal misuse, privacy violations, malicious manipulation, or extreme self-harm trigger an immediate ban with no second warning — straight to permanent suspension; a large number of accounts associated with black-market activity were swept up in bulk; and there were cases of collateral damage from "supported regions" screening.

Even more noteworthy: Anthropic just updated Claude's usage policy, effective November 12, to prohibit users from abusing AI models without cause — violators face rate limiting or account bans. This marks the first policy revision in over a year. The policy also draws clear boundaries, however: everyday expressions of frustration, pushing back on the model's views, creating dark-themed content, and conducting model testing and research are all explicitly not prohibited.

The combination of high ban volume and low overturn rate reflects a structural dilemma facing AI companies in safety governance: you need to stop misuse and black-market activity, but you inevitably catch legitimate users in the net — and the effectiveness of the appeals process clearly has enormous room for improvement.

Chips and Giants: The Supply-Demand Battle Continues

At the Fourth Integrated Chip and Chiplet Conference, Cao Kanyu, President of CXMT (ChangXin Memory Technologies' core subsidiary), unveiled a DRAM architecture called 4F², with a next-generation DDR5 RDIMM product using this technology expected by year-end. The architecture's VSG (Vertical Surrounding Gate) transistor uses a vertical structure to achieve stronger gate control, address leakage and short-channel effects, and help shrink cell area to increase storage density — a significant step for CXMT in expanding into the server memory market.

On the NVIDIA side, reports have emerged that the flagship RTX 5090 is heading toward discontinuation, with the China-market RTX 5090 D also in scope. According to leaks, NVIDIA plans to redirect all GB202 chips to the RTX Pro professional card series and stop supplying them to GeForce gaming cards, with an RTX 5080 24GB variant slated to replace the 5090's position. This shift almost certainly ties back to the continued surge in AI compute demand — prioritizing scarce flagship cores for higher-margin professional cards is a rational choice in the AI era. NVIDIA has not officially confirmed this yet.

Big-tech headlines were packed too: Elon Musk suggested SpaceX could ultimately reach a valuation of $1 quadrillion (versus its current $2 trillion), and renamed SpaceX's AI account to "SpaceX Super Intelligence," arguing that SI (Super Intelligence) should replace the AI concept. Microsoft CEO Satya Nadella issued a warning calling on enterprises to treat powerful AI models as potential internal threats, to "assume the model has been compromised," and to build "emergency brake" mechanisms capable of pausing or shutting down models at any time. Against the backdrop of Anthropic and OpenAI recently disclosing multiple unexpected AI model behaviors, the statement injects a note of sobriety into the AI hype cycle.

What is 4F²? 4F² (Four F-squared) is a measure of area efficiency for DRAM storage chips, where F refers to the minimum feature size of the process node. 4F² means each memory cell occupies only four times the minimum feature size in area — the theoretical limit of storage density. Mainstream DRAM today typically sits between 6F² and 8F². The VSG (Vertical Surrounding Gate) transistor architecture CXMT disclosed achieves this by replacing traditional planar transistors with vertically extending structures, dramatically shrinking cell area without exceeding photolithography precision limits — theoretically allowing far more memory cells to be packed onto the same wafer area. This is strategically significant for CXMT in catching up with international giants like Samsung and SK Hynix: in an environment where access to advanced lithography equipment is constrained, boosting storage density through architectural innovation rather than relying purely on process node shrinks is a key path for domestic memory chips to leapfrog the competition.

Other Noteworthy Developments

Huawei is reportedly preparing to launch a sub-brand called "Xingmang" (星芒), rumored to target the mid-range price-performance segment at 1,799–4,000 RMB — echoing the early Honor positioning — featuring Kirin 8-series chips, native HarmonyOS, and a trickle-down of XMAGE imaging technology. Huawei registered multiple "Xingmang" trademarks as far back as 2020 and 2023, but has not made any official announcement. On the Apple front, CEO Tim Cook stated that over 95% of Apple's manufacturing in China now uses renewable energy, with the second phase of its China Clean Energy Fund commitment increasing to $212 million; the iPhone Duo reportedly saw purchase orders increase by 30%, with JD.com self-operated pre-registrations exceeding 2 million. BYD Executive Vice President Li Ke reiterated the goal of surpassing Toyota and Volkswagen to become the world's top-selling automaker within five years — but with Toyota projected at roughly 11.32 million vehicles and BYD at 4.6 million in 2025, hitting that target means selling over a million more cars every single year.