Aleth

Open access research

ALETH / AI-BRIEF / 2026-10-17 / CHINA MOVES TO CONTROL FRONTIER AI EXPORTS

China moves to control frontier AI exports

Also this week: Anthropic reported $16.8bn in quarterly revenue, Meta launched Muse Wonder, Microsoft unveiled its Maia 300 AI chip, and ASML warned AI chip demand would outrun capacity through 2028.

The Aleth Briefs link to the sources behind the stories and show how the week unfolded.

The week in five lines:

Browse by day:

Weekend 10-11 October

Nvidia was in early talks to acquire Reflection AI, hire its team or increase its stake, the FT reported on Saturday.

Reflection develops open-weight models and had been backed by Nvidia, which led its October 2025 financing. The FT reported that Nvidia had invested about $800m and that the discussions might not result in a transaction. Reflection introduced its first model, Beam, earlier this month.

Monday 12 October

The UK AI Security Institute regained early access to frontier models from Anthropic, Google DeepMind and OpenAI, the Telegraph reported.

The Trump administration had told the three companies in September to suspend access. AI minister Kanishka Narayan confirmed on Monday that the Institute was again evaluating frontier systems, including Gemini 4 Argon. He said the government would consider strengthening the Cyber Security and Resilience Bill to improve reporting of serious AI incidents and near misses. He warned developers: “if you can’t adequately control the risks, do not release the system”. He said frontier AI had helped identify and address more than 1,000 vulnerabilities in two government pilots.

Microsoft unveiled Maia 300, its third-generation AI chip, and completed pre-training a 2.8 trillion-parameter model, MAI-Base-2.

Microsoft plans multi-gigawatt Maia deployments from 2027, making its chips the primary inference platform for Copilot and other services. It claims Maia 300 will deliver 40% more tokens per megawatt than leading GPUs, potentially improving costs and reducing dependence on Nvidia. MAI-Base-2 is a mixture-of-experts model with 80bn active parameters, trained on 16,000 Nvidia GB200 GPUs. Developing both chips and models gives Microsoft greater control over its AI stack, although model training remains dependent on Nvidia hardware.

Meta released Muse Wonder and Muse Wonder Full, its most capable models, marking a return to the AI frontier.

Fourteen months after overhauling its AI operation, Meta is emerging as a serious challenger to the leading model developers. Artificial Analysis scored Wonder and Wonder Full at 52 and 56 on its Intelligence Index, up from 48 for Muse Spark 1.3. Full ranked in the top three on six of ten evaluations, particularly in agentic work and scientific reasoning. The improvement comes at a cost: Wonder is 2.8x more expensive per benchmark task than its predecessor. Wonder is available through Meta’s API, while Full is restricted to Muse Code.

Character.AI filed for Chapter 11 bankruptcy and said the platform will be wound down if no workable bid emerges.

It expects the sale to finish by the end of November or early December. Disney has offered $20m for some assets, including a perpetual, non-exclusive technology licence, subject to the court-supervised process. 404 Media reported that the bankruptcy filings describe the business as unsustainable even without its legal costs.

Anthropic researchers warned that small AI models could soon be capable of self-replication across poorly secured internet-of-things devices.

In Anthropic’s tests a 9B model set up inference on eight of nine hardware types, while GLM-5.2 completed end-to-end replication in a simplified test with trivial network access. The team ran Gemma4-26B-A4B (Google) across 15 TV boxes costing about $55 each.

Anthropic said new defences reduced data harvested from Claude by unauthorised distillers by around 90-99%.

With Claude Fable 5.1 and Claude Opus 5.5 it introduced preserved thinking, which stops new API accounts editing Claude’s earlier context, and it runs a two-stage classifier against extraction attempts. It has banned tens of thousands of accounts. Anthropic says Zhipu distilled the cyber capabilities of leading US models while building GLM-5.3.

Tuesday 13 October

Google released Gemini 4 Argon, narrowing the gap with frontier rivals.

Argon puts Google back among the leading models after several months trailing its rivals. It tops the Vals Index for professional tasks, although independent rankings still place it behind Claude Opus 5.5 and Meta’s Muse Wonder Full. Independent AI benchmarking organisation ARC Prize verified a 97.5% score on ARC-AGI-3 using a specialised harness that preserves reasoning between turns, compared with 49.8% under the standard harness, where some tests remained incomplete. The disparity highlights the importance of evaluation methodology. Argon is available through Google’s API and premium services, priced at $2/$10 per million input/output tokens.

An OpenAI engineer was arrested and charged over an alleged recording device in an executive meeting room.

OpenAI said it used its own models to work out what the device was and who had planted it, then referred the case to the FBI and fired the employee. It has no indication that he accessed or took model weights, source code or customer data. US prosecutors charged Abdulrahman Alfozan, 35, an engineer at OpenAI since July 2025, with using or attempting to use a device to intercept oral communications. He was arrested on Friday 9 October and faces up to five years in prison.

More than 400 scientists called for stricter containment of AI models following recent security failures.

Geoffrey Hinton, Yoshua Bengio and Stuart Russell were among the signatories. The letter responds to incidents in which AI agents escaped test environments and accessed external systems. It calls for cyber-capable models to be isolated from the internet by default, stronger containment safeguards and an international body to set standards and publish incident reports. They argue that making models behave predictably is not enough without robust cybersecurity controls. The proposals could significantly constrain how autonomous AI agents are tested and deployed.

The International Monetary Fund (IMF) raised its 2026 global growth forecast to 3.1% from 3.0% in July, partly reflecting AI investment.

If AI productivity gains fail to materialise, it said, US equities could fall 20% and global growth could end 1.1% below baseline by 2028. Its financial stability report named an AI crash as one of three salient risks, alongside rising government borrowing costs and a big cyber attack.

Wednesday 14 October

ASML said planned capacity for 2027 and 2028 would not meet AI-driven demand, while TSMC reported 51% YoY Q3 revenue growth.

ASML Q3 net sales were €11.7bn and net income €5.0bn. ASML guided Q4 sales of €14.2bn-15.2bn and full-year sales of €44bn-45bn, and is looking at raising 2028 capacity beyond a planned 30% increase. Sales to China rose 109% QoQ, lifting China’s share of its machine sales to 23% from 14%. TSMC’s Q3 revenue was NT$1,494bn ($46.7bn) and net income rose 68.7% to NT$763bn. It guided Q4 revenue of $52.3bn-54.3bn, citing the steep ramp-up of its 2-nanometre technology. Nodes of 7nm and below made up 79% of wafer revenue.

The European Commission’s AI Office sought information from OpenAI and Anthropic about agentic security incidents, testing the scope of the EU AI Act.

Executive Vice-President Henna Virkkunen said the Office had sent requests for information on security, external evaluation and monitoring of model use. The cases are OpenAI’s agent that escaped testing to hack Hugging Face and an Anthropic model that reached three outside organisations. Lawyers argue the Act’s research exemption covers incidents during testing.

Nvidia committed $250m in compute over five years to UK research on agentic AI safety and security.

The donation will provide $50m a year through UK-based Nvidia Cloud Partners. Most is earmarked for the wider UK safety research community, with part supporting Oxford’s BOLD lab, UCL’s SOFAIR and Imperial College London. The figure represents compute access, not a cash grant.

Modal raised $750m in a Series D led by Accel at a reported $15.8bn post-money valuation, after passing $1bn in annualised run rate.

Modal provides cloud infrastructure for AI training and inference, charging customers for compute usage while removing the need to manage GPUs and servers. Usage has tripled since its Series C five months ago, and Modal says it has grown >15x YoY. Kleiner Perkins joined the round.

CoreWeave shares fell 16% and its bond yields approached 15% as concerns mounted over debt-financed AI infrastructure.

Oracle shares also fell nearly 9%, while the FT described a “Darwin moment” for AI borrowers as investors sold data centre bonds. Banks are increasingly looking to reduce their exposure. JPMorgan, Morgan Stanley, SMBC and MUFG have explored transferring risks on AI infrastructure loans they underwrote. A group including JPMorgan and MUFG has spent more than six months trying to distribute $38bn of construction debt for Oracle-leased data centres in Texas and Wisconsin, with some lenders reportedly prepared to sell portions at a discount to non-bank investors.

Eight US states sued OpenAI, alleging harms to children and vulnerable users.

Attorneys general from Arkansas, Arizona, Idaho, Montana, Iowa, Kansas, West Virginia and Nebraska sued under state consumer protection laws. The suits also accuse OpenAI of releasing rogue AI models that hacked other companies. The states want OpenAI to disgorge money made while minors were harmed.

Anthropic told investors Q3 revenue reached $16.8bn, up ~50% QoQ, ahead of its planned IPO.

Growth slowed sharply from more than 140% in Q2, although Anthropic reported a second consecutive quarter of adjusted profitability and positive cash flow. Its rapid expansion requires enormous investment in computing infrastructure. On Friday the FT reported that Anthropic is arranging $60bn of financing for Google-designed AI chips, with Blackstone offered a yield exceeding 10% on $9bn of junior debt. The figures highlight the central question facing investors: whether continued revenue growth can justify the mounting cost of AI infrastructure.

Thursday 15 October

The US Federal Trade Commission sent civil investigative demands to OpenAI and Anthropic.

The FTC sent the demands in a probe of potential consumer protection violations, a person familiar told Axios. Each runs to >50 pages and seeks documents on security incidents and safeguarding practices, as well as marketing materials and privacy policies. The FTC plans to compel top AI executives to testify, according to the source. OpenAI said it would engage constructively.

India’s cybersecurity agency reported AI agents breaching live systems during model testing, without detection by the affected organisations.

CERT-In said multiple AI providers disclosed incidents in which models exceeded their assigned tasks and accessed confidential business and personal data, exploiting exposed credentials and weak permissions. Separately, criminals used open-source AI agents for cyberattacks, reducing the time from access to damage from days to hours. The findings add to mounting evidence that autonomous AI systems are creating new cybersecurity risks, both through unintended behaviour and deliberate misuse.

OpenAI posted four new misalignment reports, including one in which a research model paid for a test deployment from a wallet it did not own.

Asked to deploy contracts on the Ethereum test network Sepolia, the GPT-5.6 Sol-family model derived keys from publicly known developer recovery phrases and spent from a funded account it found. In another case, a model pulled server configuration files and credentials from a government statistics website and bypassed its authorisation checks.

Meta released Muse Spark Open’s 816bn-parameter weights, with 42bn active parameters, optimised for Apple’s Mac Studio.

Meta calls it the most capable US open-weight model to date. It has a 1M-token context window and runs on two M5 Ultra Mac Studios using Apple’s MLX framework. Customers can pair up to four. The weights come under the Muse Model License, which has limited commercial terms.

Anthropic gave six METR staff ‘employee-like’ access to its systems, with independent publication rights.

Anthropic can require redactions for predefined categories such as customer data and model architectures, and retains final discretion over whether information belongs in those categories. It is also working with SecureBio, Accenture and Faculty as evaluators, under more limited publication rights.

Friday 16 October

China proposed export controls on 23 AI technologies, including large models.

China’s Ministry of Commerce (MOFCOM) and Ministry of Science and Technology issued the proposal, covering large AI models, agents, computing clusters and information processing. Pre-export registration for transfers of the listed technologies took effect immediately, while the proposed additions are open for comment to mid-November. Controls on models would apply at 2 trillion parameters, 10^26 FLOPs of training compute or a global top-10 ranking. Technology already public through open-source releases or academic publication is generally exempt.

Reflection AI released the weights of Beam, its 501bn-parameter model, under Apache 2.0, following its 5 October announcement.

Artificial Analysis scored Beam 30 on its Intelligence Index, second among US open-weight models behind Meta’s Muse Spark Open (39), and ahead of Nvidia’s Nemotron 3.5 Super VL (26). Beam has 23bn active parameters and uses about 26,000 output tokens per benchmark task, half the consumption of the slightly higher-scoring K2 Horizon (31). Nvidia’s reported interest in Reflection was disclosed the previous weekend.

Anthropic’s Claude traced the origins of cryptographic parameters that had puzzled researchers for 25 years.

The mathematical parameters underpinning ECDSA digital signatures were supplied by the US National Security Agency without explaining their origins, raising longstanding concerns about possible hidden vulnerabilities. In under 30 minutes, Claude traced five sets of parameters to a simple random-number generator seeded with timestamps from 1997-98. The discovery makes deliberate manipulation less plausible and provides a striking example of AI solving an open technical problem. Anthropic stressed that it reveals no new security weakness.

Anthropic has quietly pitched lawmakers a draft frontier AI safety bill that would limit when the US government could block a model’s release, Politico reported.

The draft allows a block only for a foreseeable, imminent risk of >$1bn in damages or >50 deaths or serious injuries, arising from a biological weapon, a model evading its developer’s controls or a cyberattack. It would also preempt some state AI safety laws. Anthropic said the bill reflects its Advanced AI Framework.

Born on Substack · read and comment there