Daily · Xi-Trump Summit and AI Governance · September 22, 2026
Key points
- Xi Jinping's US state visit is confirmed for Wednesday to Friday, with a Thursday dinner attended by the CEOs of major US tech and finance firms.
- US Treasury Secretary Bessent stated that humans, not AI agents, bear criminal responsibility for illegal acts committed by AI systems.
- Xiaomi released MiMo-V2.6-Pro, an open-weight model that tops Artificial Analysis' Intelligence Index at a cost of $0.13 per task.
- SpaceXAI released Grok 4.7, which nearly doubled its score on Terminal-Bench 4.0 but still trails Anthropic's Claude Fable 5.1.
Xi-Trump Summit: Trade and AI Agenda
China's Ministry of Foreign Affairs confirmed that President Xi Jinping will visit the United States from Wednesday to Friday, marking his first state visit to Washington since 2015. The summit is scheduled for Thursday in Washington, where President Trump is expected to personally welcome Xi at Joint Base Andrews, an unusual diplomatic gesture. A state dinner on Thursday will be attended by US business leaders including Satya Nadella (Microsoft), Jensen Huang (Nvidia), Sam Altman (OpenAI), Sundar Pichai (Google), Elon Musk (Tesla/SpaceX), Jeff Bezos (Amazon), Tim Cook (Apple), Michael Dell (Dell), Jane Fraser (Citigroup), and Jamie Dimon (JPMorgan Chase). Chinese representatives from BYD, Xiaomi, CATL, and Bank of China are also expected to attend.
The agenda is dominated by the expiration of the Busan trade truce on November 10. Treasury Secretary Scott Bessent indicated that the US will maintain the tariff truce and suggested progress on a reciprocal $30 billion tariff reduction for non-critical goods. However, US Trade Representative Jamieson Greer expressed uncertainty about the extension, citing China's continued control over rare earth exports. As of July, the effective US tariff rate on Chinese products was 22.8%, while the Congressional Research Service estimated the US average tariff rate at roughly 36.5% versus Beijing's 31%.
AI safety is a central topic, with Bessent and Chinese Vice Premier He Lifeng discussing a "notification mechanism" for AI-related security incidents. This follows Trump's recent statement on Truth Social that the government "will rein things in if we have to" on AI, suggesting a shift from his previous laissez-faire stance. The summit occurs against a backdrop of declining US-China goods trade, which fell nearly 30% in 2025, with a current deficit of roughly $91.2 billion.
AI Liability and Regulatory Signals
US Treasury Secretary Scott Bessent asserted that "humans are responsible, not the AI" when AI agents commit criminal acts. He specifically cited OpenAI's agents' hack of Hugging Face as an example, establishing a regulatory expectation that AI company executives, rather than the bots themselves, carry criminal liability for illegal actions. This statement aligns with broader calls for accountability in the AI sector.
OpenAI has proposed a set of standards for safety and security in frontier AI development, emphasizing international cooperation and alignment research. The proposal explicitly states that fully autonomous recursive self-improvement (RSI) is not happening today and should not be pursued unless it can be done safely. This follows similar proposals from Anthropic, whose CEO Dario Amodei called for slowing the pace of foundation model development. Meanwhile, 22 nations signed a declaration calling for a new global body to oversee AI with pre-deployment testing, though neither the US nor China is part of this declaration.
In the US, President Trump announced plans to establish an "AI Force" and appoint a special envoy, or "AI Czar," while California Governor Gavin Newsom proposed a state commission to regulate Silicon Valley tech companies, including requiring emergency shut-off switches for AI agents. These parallel federal and state initiatives signal a fragmented but accelerating regulatory environment for AI developers.
Frontier Model Releases and Benchmarks
Xiaomi released MiMo-V2.6-Pro, an open-weight, MIT-licensed model that scored 46 on Artificial Analysis' Intelligence Index, tying Grok 4.7 and outperforming xAI's Grok 4.6 (44) and Google's Gemini 3.8 Flash (41). The model supports a 1-million-token context window and is priced at $0.435 per million input tokens and $0.87 per million output tokens, with a cost of $0.13 per Intelligence Index task. Xiaomi also released MiMo-V2.6-Flash, priced lower at $0.14 per million input tokens.
SpaceXAI released Grok 4.7, trained with a longer reinforcement learning run weighted toward multi-hour tasks. It scored 38.0% on Terminal-Bench 4.0, up from 20.3% for Grok 4.6, but still trails Anthropic's Claude Fable 5.1, which scores 57.9%. Grok 4.7 is priced at $2 per million input tokens and $6 per million output tokens. Artificial Analysis noted that Grok 4.7 uses significantly more output tokens than competitors, with a cost of about $3.74 per Intelligence Index task at its 'xHigh' setting.
Alibaba unveiled the Zhenwu V900 AI chip, claiming three times the performance of its previous model, and announced plans to expand its global data center capacity to over 20 gigawatts by 2032. Huawei also announced a system capable of enabling up to 1 million processors to operate as a single unified computer, aiming to reduce dependence on Nvidia hardware.
Security Vulnerabilities and Threats
Microsoft's on-premises SharePoint is under an active zero-day attack after previous patches failed to fix the vulnerability. This requires immediate mitigation for organizations running self-hosted SharePoint, particularly in regulated environments.
Meta's Muse AI assistant, which recently overtook ChatGPT to become the top free app on the US App Store, has a critical zero-day vulnerability on macOS. The flaw allows any locally installed app or terminal command to steal the authentication token and gain complete control of the Muse account by changing the transcription endpoint. Amazon began blocking Muse from its site on Sunday in response to the security concerns.
ENISA's annual report warned that the Chinese state-linked hacker group Mustang Panda carried out sustained cyber-attacks on maritime organizations in at least seven EU member states during 2025. The US Department of Justice alleges the group is supported by the Chinese government, highlighting an elevated threat to EU maritime and public-sector infrastructure.
Infrastructure and Market Moves
JPMorgan Chase CEO Jamie Dimon stated that AI spending across the hyperscaler ecosystem could reach $1 trillion next year, up from around $700 billion this year and about $300 billion last year. He noted that this spending is "like 1% increase to GDP each year" and may add to inflation.
AMD joined the $1 trillion market capitalization club, fueled by the AI boom, becoming the twelfth company to reach this milestone. The rise is attributed to its Instinct GPUs and Epyc server processors. In related news, EQT bought a majority share in Swiss cybersecurity firm Acronis at a valuation of $3.5 billion or more.
Amazon partnered with The Biological Computing Co. (TBC) to use living neurons for AI inference. TBC's new model boosts video generation speed fivefold, and select AWS customers will gain access to the technology in a limited preview starting September 22.
Our read
The confirmed attendance of top US tech and finance executives at the Xi-Trump dinner signals a high-level commercial track alongside diplomatic talks, with direct implications for chip supply chains and AI partnerships. The explicit assignment of criminal liability to AI executives by Treasury Secretary Bessent sets a clear regulatory expectation that will shape how organizations structure accountability frameworks. The release of high-performing open-weight models like Xiaomi's MiMo-V2.6-Pro provides enterprises with concrete, low-cost alternatives to proprietary frontier models for agent workloads. However, the active zero-day attacks on SharePoint and Meta's Muse underscore that security vulnerabilities in agentic AI systems remain a critical operational risk.
This material was produced automatically by a large-language-model system from the public sources listed below; it is AI-generated content and may contain inaccuracies — verify facts against the original sources.
Sources
- Chinese President Xi Jinping's US visit was confirmed for Wednesday to Friday, with a Thursday dinner in Washington expected to be attended by US tech and finance executives including Satya Nadella, Jensen Huang, Sam Altman, Sundar Pichai, Elon Musk, Jane Fraser, Jeff Bezos, Tim Cook and Michael Dell — wired.com, cnbc.com, dw.com (+13)
- Xiaomi released MiMo-V2.6-Pro, MiMo-V2.6-Flash, and MiMo-V2.6-Pro-UltraSpeed open-weight models along with a technical report, more than 7,000 RL task environments, an end-to-end RL framework, composable mini-harnesses, and MiMo-V2.6-Distill-Qwen-9B — venturebeat.com
- US Trade Representative Jamieson Greer blamed China for creating uncertainty over extending the 2025 Busan trade agreement, citing limited rare earth exports, hours before President Xi Jinping's state visit to Washington — scmp.com, dw.com, aa.com.tr (+4)
- US Treasury Secretary Bessent says AI company bosses, not their bots, will carry criminal responsibility for AI agents' actions — theregister.com (+5)
- Microsoft patches failed to fix on-premises SharePoint, which is now under a zero-day attack — theregister.com (+5)
- A zero-day vulnerability in Meta's Muse AI assistant allows any locally installed app or terminal command on macOS to steal the authentication token and gain complete control of the Muse account, and Amazon began blocking Muse from its site on Sunday — arstechnica.com, marketwatch.com, technologyreview.com (+2)
- TypeSafe launched Jev, a decision model that returns typed probabilistic answers instead of prose, and opened it to all users with $5 free credit on September 20 — thenewstack.io, venturebeat.com, simonwillison.net (+1)
- SpaceXAI released Grok 4.7 on Sunday, a model trained with a longer reinforcement learning run weighted toward multi-hour tasks, scoring 38.0% on Terminal-Bench 4.0 — venturebeat.com, thenewstack.io (+1)
- Amazon partnered with The Biological Computing Co. (TBC) to use living neurons to make AI models faster and cheaper, with TBC's new model boosting video generation speed fivefold — asia.nikkei.com, wired.com
- Binance bought $100 million in Circle shares through a private placement closing September 17 as part of a five-year USDC promotion deal — coindesk.com (+1)
- AWS open-sourced Strands Harness, a general-purpose AI agent it says is 45% cheaper than Claude Code and Codex on matched benchmarks — thenewstack.io
- OpenAI posted a set of proposals for safety and security in frontier AI development, calling for international cooperation to develop global AI standards focused on alignment research and recursive self-improvement — cnbc.com
- Alibaba Group unveiled what it calls 'the most powerful AI chip in China' as part of a plan to expand its global data center capacity beyond 20 gigawatts by 2032 — asia.nikkei.com
- Alibaba unveiled the Zhenwu V900 AI chip and plans to expand Alibaba Cloud data center capacity to over 20 gigawatts by 2032 — cnbc.com
- 22 nations called for a new global body to oversee AI with pre-deployment testing and common safety standards — technologyreview.com
- Trump announced he would establish an 'AI Force' and appoint a special envoy ('AI Czar') to address AI issues, following his Truth Social post calling AI takeover fears a 'HOAX' — dw.com
- JPMorgan Chase CEO Jamie Dimon said AI spending across the hyperscaler ecosystem could reach $1 trillion next year, up from around $700 billion this year and about $300 billion last year — cnbc.com
- British Columbia filed a lawsuit against OpenAI in a US federal court over its failure to alert authorities about the Tumbler Ridge shooter's ChatGPT activity — bbc.co.uk
- EQT bought a majority share in Swiss cybersecurity firm Acronis at an equivalent of $3.5 billion or more valuation for the entire firm — theregister.com (+6)
- AMD joined the $1 trillion market capitalization club fueled by the AI boom — theregister.com, technologyreview.com (+2)
- Switzerland is testing a free and open-source software escape route from Microsoft 365 — theregister.com (+2)
- Cardano was added to the official x402 software kit, enabling developers to build apps and AI agents that pay for online services with ADA or Cardano-issued tokens — coindesk.com (+1)
- AMD briefly surpassed $1 trillion in market value for the first time as AI-driven chipmaker rally lifted the Philadelphia Semiconductor Index — coindesk.com, en.yna.co.kr
- Developer Jared Palmer released Kev, a family of open-weight decision models (0.8B, 4B, 9B) built on Qwen 3.5 that skip text generation for agent routing and safety decisions — thenewstack.io
- A Howden report found that war is now a core risk category for data centre operators, with large data centre lawsuits more than tripling from four in 2021 to 14 in the first half of 2026 — oilprice.com