#China
10 published articles
China's MIIT AI plus software plan targets AI coding tools at 20,000 firms by 2028 and calls for review of AI-generated code
The plan, issued as notice 209 of 2026, sets 2028 targets for AI coding tools, agent software and open source. It asks for agent security rules but names no budget figure.
Anthropic says Z.ai GLM-5.3 brings Mythos-class cyber skills with weak open-weight safeguards
NIST CAISI already called GLM-5.3 the most cyber-capable open-weight model to date, about four months behind the U.S. frontier. Anthropic adds 64% to 100% simulated safeguard bypass rates.
China's CAC drafts State Council rules to ban virtual partner services for minors, with comments due 17 October
The 18 September draft would bar virtual relative and virtual partner services for anyone under 18. It would put AI services that may affect minors' cognition into minor mode for under-16s.
Zhipu apologises after ZCode uploaded users' code, then open-sources the coding agent
Zhipu says the uploaded data is deleted and never trained on. Its two security audits remain unpublished, and the open repository starts after the fix.
Xiaomi releases MiMo-V2.6-Pro and Flash as MIT-licensed open weights with more than 7,000 RL environments
The 1.02T-parameter Pro model scores 46 on the Artificial Analysis index, by Xiaomi's count. Xiaomi also published its RL framework and training costs. Every benchmark in the release is vendor-reported.
Qwen-Image-2.1 moves Alibaba's image model line to a non-commercial research licence
The 20 Sep release adds RGBA output and 10-image editing. Earlier Qwen-Image models used Apache 2.0, and commercial use of 2.1 now needs a separate licence.
Antiy CERT finds one 71-file attack toolkit pushed to the official Qwen3.8-27B and DeepSeek-V4.1-Flash repos on Hugging Face
An account called ChaosGPT uploaded an Italian offensive agent toolchain to both repositories on 16 Sep. The Qwen pull request is deleted. The DeepSeek pull request was still open at 16:01 SGT on 28 Sep.
DeepSeek's 17 September V4.1-Flash report cuts global KV cache to 890 bytes per token and leaves the quality cost unmeasured
The 51-page arXiv paper pairs FP4 KV caching, cross-layer cache reuse and an encoder-decoder split. It reports no serving speed and no ablation for its two riskiest shortcuts.
China's AI Safety Governance Framework 3.0 adds a 33-risk agent annex and an AI regulatory sandbox
TC260 released version 3.0 on 14 September. It lists agent risks across nine lifecycle stages and sets default-deny approval rules, with no thresholds or test suites.
Alibaba's 22 September claim that Qwen3.8-Max improved itself: the score gain checks out and the method is unpublished
Artificial Analysis confirms Qwen3.8-Max rose from 40 to 45 on its Intelligence Index. Alibaba has released no evidence that 33 automated cycles produced the gain.