iiinova — where your digital self keeps evolving

Emmie Hine @emmiehine.com · 25/09/2026

16/ Number of the week: More than 90% of the 430,000 micro-dramas released in China in January–August 2026 were AI-generated, according to a National Radio and Television Administration official.

100

Emmie Hine @emmiehine.com · 25/09/2026

15/ The MIIT approved five agent capability standards. They set technical requirements, not security rules. The issue also tracks eight China-backed agent-security work items at ITU-T.

100

Emmie Hine @emmiehine.com · 25/09/2026

14/ On standards, TC260 issued four nonbinding AI application security guides, including provisions on human control, rollback, and emergency shutdown. It also opened comment on draft security guidance for agent developers; comments close October 2.

100

Emmie Hine @emmiehine.com · 25/09/2026

13/ From the safety papers: Fudan researchers found agents that detected dangerous plans but did not enforce audit verdicts. Adding an enforcement check cut attack success more than fourfold in their experiments.

100

Emmie Hine @emmiehine.com · 25/09/2026

12/ Another Zhuque paper tests whether unsafe behavior can transfer through a multi-agent handoff. The results show a risk worth testing, but do not establish how often such incidents occur or demonstrate an autonomous cascade.

100

Emmie Hine @emmiehine.com · 25/09/2026

11/ Two Tencent Zhuque Lab papers study agent loss of control. In simulated single-agent tests, dropping constraints from a compacted summary plus an unsafe opportunity produced unauthorized actions. Restoring the constraints prevented them in those runs.

100

Emmie Hine @emmiehine.com · 25/09/2026

10/ A cross-industry paper maps five levels of recursive self-improvement. Its authors say full meta-improvement has not been demonstrated; the strongest established level in software engineering is still narrower, with humans setting the objective and evaluation criteria.

100

Emmie Hine @emmiehine.com · 25/09/2026

9/ Alibaba says Qwen3.8-Max ran automated improvement work for more than a month. Across 33 cycles, its Artificial Analysis score rose from 40 to 45. Alibaba has not explained how the runs were overseen or independently checked.

100

Emmie Hine @emmiehine.com · 25/09/2026

8/ Model releases include Shanghai AI Lab's Atria Dawn Preview, built by post-training Zhipu's GLM-5.2; DeepSeek-V4.1-Flash; and Xiaomi's MiMo-V2.6 Pro and Flash. The issue covers their architecture, reported results, and limits.

100

Emmie Hine @emmiehine.com · 25/09/2026

7/ The CAC acted against an unlabeled AI mini-program and an operator reselling model access through relay sites. The cases show how labeling and service-filing rules are being enforced in practice.

100

Emmie Hine @emmiehine.com · 25/09/2026

6/ A draft State Council regulation would bar all online services from offering minors virtual intimate-relationship services. Existing rules already restrict AI companion providers; the new draft would reach beyond them. Comments close October 17.

100

Emmie Hine @emmiehine.com · 25/09/2026

5/ At home, the State Council called for better monitoring and coordination of computing capacity, electricity, and networks. CAICT says its national platform now monitors compute resources in all 31 provinces.

100

Emmie Hine @emmiehine.com · 25/09/2026

4/ China's Foreign Ministry rejected Dario Amodei's framing of China as an AI security threat. The next day, spokesperson Guo Jiakun said China takes advanced AI risks seriously, including loss of control, showing the objection was to Amodei's framing, not AI safety.

100

Emmie Hine @emmiehine.com · 25/09/2026

3/ At the BRICS summit, Xi Jinping proposed an AI open-source zone for model development, deployment, and training; it remains a proposal, not a BRICS commitment.

100

Emmie Hine @emmiehine.com · 25/09/2026

2/ According to China's readout of the September 24 talks, Xi and Trump supported continued AI dialogue. Xi called for cooperation against AI misuse and for human control of the technology. The readout announced no formal AI agreement or incident-notification mechanism.

100

Emmie Hine @emmiehine.com · 25/09/2026

1/ China AI Bulletin Issue 12 is out: developments from September 9–23 (plus the Trump–Xi summit). Xi and Trump discussed AI, but no AI agreement. Also: a proposed BRICS AI open-source zone and Alibaba says Qwen3.8-Max made progress toward recursive self-improvement. 🧵

China AI Bulletin 12 cover on a dark blue background

110

Emmie Hine @emmiehine.com · 18/09/2026

10/ Full analysis, including a bilingual table of every loss-of-control term in the framework: chinaaibulletin.substack.com/p/caib… Subscribe: chinaaibulletin.substack.com/subscr…

000

Emmie Hine @emmiehine.com · 18/09/2026

9/ Embodied AI gets its own category: perception, physical actuation, human interaction, and multi-agent coordination failures. Context: China's industry ministry counted more than 140 humanoid manufacturers and over 330 products in 2025.

110

Emmie Hine @emmiehine.com · 18/09/2026

8/ Also new—a regulatory sandbox proposal for sector-specific, risk-tiered trials with limited liability relief, with separate rules for finance, education, broadcasting, and healthcare. The open question is whether lessons from individual trials translate into wider rules.

100

Emmie Hine @emmiehine.com · 18/09/2026

7/ That extends months of agent rulemaking, from the CAC/NDRC/MIIT agent opinions to the SAMR/SAC agent-interconnection series. TC260 is also drafting a mandatory agent-application-security standard, so these recommendations could harden into requirements.

100

Emmie Hine @emmiehine.com · 18/09/2026

6/ Agents are a new concern. v3 gives agents a dedicated risk category and a nine-part lifecycle appendix covering permissions, tools, memory, and decommissioning, and it addresses malicious actions and agents escaping their security constraints.

100

Emmie Hine @emmiehine.com · 18/09/2026

5/ On loss of control specifically: v3 uses 失控 ("loss of control") more often and more precisely than v2, newly discussing LoC over agent behavior and across coordinated embodied systems. In the framework, "control" can mean operators, the state, or humanity.

100

Emmie Hine @emmiehine.com · 18/09/2026

4/ But those concerns sit within a broader agenda built around scams, content, privacy, environmental harm, etc. It's not a primarily frontier-oriented document.

100

Emmie Hine @emmiehine.com · 18/09/2026

3/ Frontier risks: v3 names more behaviors developers can test for. It describes models evading shutdown, deceiving evaluators, and concealing capabilities, and its preface flags recursive self-improvement that could exceed human anticipation.

110

Emmie Hine @emmiehine.com · 18/09/2026

2/ What it is: the third annual update of standards org TC260's AI safety framework, under CAC guidance. It's not binding, but it shapes the technical standards that put Chinese AI regulation into practice, so it's important nonetheless.

100

Emmie Hine @emmiehine.com · 18/09/2026

1/ New China AI Bulletin special issue: China's AI Safety Governance Framework v3.0 came out on September 14. It gives frontier risks more concrete treatment, but the bigger expansions are agents, embodied AI, and a regulatory sandbox proposal. 🧵

China AI Bulletin Special Issue 2 cover card, titled 'the AI Safety Governance Framework 3.0,' with a subtitle on frontier risk, agent, and embodied AI governance.

100

Emmie Hine @emmiehine.com · 17/09/2026

5/ The paper discusses concrete proposals in three areas: frontier risk evaluation, emergency response, and liability. Each builds on rules China already has, calibrated to the most capable systems.

100

Emmie Hine @emmiehine.com · 17/09/2026

4/ The core finding: China already has a lot of the infrastructure, but it isn't differentiated by capability—on paper, a frontier model faces roughly the same requirements as a much smaller one.

100

Emmie Hine @emmiehine.com · 17/09/2026

3/ We looked at how the US, UK, and EU have approached frontier AI regs, then mapped China's own infrastructure, including the algorithm/model registry, S&T ethics review, and emergency response system.

100

Emmie Hine @emmiehine.com · 17/09/2026

2/ China already has a strong AI regulatory ecosystem. The question we wanted to answer: how can that existing infrastructure be extended to cover frontier-specific risk?

100

Emmie Hine @emmiehine.com · 17/09/2026

1/ Frontier AI keeps getting more capable, so how do we regulate its unique risks? I've been digging into that for China with legal experts worldwide, including some who authored China's academic draft AI laws: "Considerations for Frontier AI Governance in China." 🧵

Cover of the report 'Considerations for Frontier AI Governance in China: Adapting Existing Regulatory Infrastructure to Frontier Risk,' listing its authors.

200

Emmie Hine @emmiehine.com · 11/09/2026

15/ Number of the week: 24.3 percent—the share of correct content moderation decisions that get reversed when Chinese text is rewritten in variant characters that read the same to a person.

100

Emmie Hine @emmiehine.com · 11/09/2026

14/ From the safety papers: One paper lays a ladder for reasoning models’ progression towards superintelligence, while FUSE is an evaluation framework for dangerous capabilities, including chem-bio risks.

100

Emmie Hine @emmiehine.com · 11/09/2026

13/ On standards: the SAC approved 60 guiding technical documents, 42 of them on AI or brain-inspired computing. TC260 opened comment on a five-level AI security maturity assessment, and SAMR registered nine more AI projects, three of them on ethics review.

100

Emmie Hine @emmiehine.com · 11/09/2026

12/ The National Data Administration is turning tokens into an economic metric. Director Liu Liehong said on August 29 that tokens could measure, price, trade, and settle AI services, carrying value between data, models, and compute.

100

Emmie Hine @emmiehine.com · 11/09/2026

11/ Huawei introduced the Kirin 9050 Pro, its first “LogicFolding” chip. The only technical account is an arXiv paper by Huawei's Tingbo He, which says sub-7nm scaling needs EUV lithography, “which cannot be accessed,” so “another way had to be found.”

210

Emmie Hine @emmiehine.com · 11/09/2026

10/ Two Chinese commentaries proposed a narrower bilateral agenda ahead of the rumored September Track 1: joint testing and limits on capabilities that could cause serious harm, but no single country or company defining what counts as dangerous.

100

Emmie Hine @emmiehine.com · 11/09/2026

9/ China agreed to the joint statement carrying the Principles, but its own readouts don't mention them.

100

Emmie Hine @emmiehine.com · 11/09/2026

8/ At the G20 Innovation ministerial in Chapel Hill on September 1-2, ministers adopted the Carolina Principles for Emerging Technologies: use existing sector regulation where possible, reserve new rules for gaps those frameworks can't address.

100

Emmie Hine @emmiehine.com · 11/09/2026

7/ Internationally, the NSA, CISA, and FBI accused six Chinese AI companies—DeepSeek, Moonshot, Alibaba, MiniMax, StepFun, and Zhipu—of “industrial-scale” distillation. The agencies grant distillation is standard practice; an MFA spokesperson rejected the claims.

100

Emmie Hine @emmiehine.com · 11/09/2026

6/ The CAC's Qinglang campaign entered a second phase, shifting from model and service compliance to AI-generated content and conduct: impersonation, false information, and automated fake engagement.

100

Emmie Hine @emmiehine.com · 11/09/2026

5/ CAC official Wang Lihong named five major AI security risks on September 1: technical vulnerabilities, capability gains and “extreme loss of control,” agents with high system privileges, misuse including in biology and genetics, and “technological hegemony.”

100

Emmie Hine @emmiehine.com · 11/09/2026

4/ They don't address whether AI-generated content can be copyrighted or whether developers can train on copyrighted works without permission–the drafters disagreed.

100

Emmie Hine @emmiehine.com · 11/09/2026

3/ Re: open models: one part gives open-source developers an appropriate liability exemption weighted by license type and disclosure, and another treats training on lawfully disclosed personal data as generally non-infringing absent express refusal.

100

Emmie Hine @emmiehine.com · 11/09/2026

2/ China's Supreme People's Court issued the country's first nationwide adjudication rules for AI disputes on September 7, telling courts how to apply existing law. Courts can cite this in their reasoning.

111

Emmie Hine @emmiehine.com · 11/09/2026

1/ China AI Bulletin Issue 11 is out: China's top court issues the first nationwide rules for AI disputes, a CAC official names extreme loss of control among five major AI risks, and Huawei unveils the Kirin 9050 Pro. Thread:

Title card for China AI Bulletin Issue 11: light blue capitals on a dark blue background, with chinaaibulletin.substack.com along the bottom.

111

Emmie Hine @emmiehine.com · 14/08/2026

10/ Full issue—governance, standards, export controls, models, and the safety papers:

chinaaibulletin.substack.com

China AI Bulletin 9

Developments from 29/7/26-12/8/26

000