Aleth

Open access research

ALETH / AI-BRIEF / 2026-09-12 / NAVIER-STOKES BREAKTHROUGH SPARKS CREDIT FIGHT

Navier-Stokes breakthrough sparks credit fight

Also this week: 25 Fields medallists pushed back on AI maths, Russian spies rebuilt malware with Claude, Samsung led a €3bn bet on Mistral and Meta launched Muse.

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

Browse by day:

Weekend (5-6 September)

OpenAI said it has no standard for reporting misalignment and is writing one.

The admission came on Saturday, over the “wiki incident” in which its agents wrote to several internet sites. OpenAI disclosed the Hugging Face breach the next day under its security playbook. The wiki case looked like ordinary misalignment, and nothing routed it anywhere. A framework is promised within weeks.

Anthropic’s $1.5bn book settlement sparked a fight over who collects.

Authors of nearly 500,000 titles are due $3,000 per pirated work, split 50-50 with the publisher where the book is still in print. Writers report publishers claiming titles whose rights reverted years ago, and claiming 100% where they are owed half. Literary agencies filed claims too, though agents are not rightsholders.

OpenAI’s research organisation now runs 3.1 agent-workdays for every human workday.

The 90th percentile researcher spends >$7,000 of tokens a day. OpenAI says it has reached the “automated research intern” threshold it set last autumn, a system doing well-defined research tasks under human direction, and aims at a fully automated AI researcher by March 2028. Chief scientist Jakub Pachocki added the caveat: “no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.” He expects voluntary slowdowns to spread until shared safety bars exist.

Monday 7 September

Insilico’s AI-designed drug lowered biological age estimates in a small trial.

Insilico Medicine analysed blood samples from 42 patients treated with rentosertib, its AI-designed drug for idiopathic pulmonary fibrosis. Six independently developed proteomic ageing clocks all indicated a reduction in predicted biological age. The strongest signal was seen after four weeks at 30 mg twice daily, with estimates falling by around three to four years. Insilico used AI both to identify the drug’s target, TNIK, and to design the molecule. The data are encouraging, but do not establish whether the changes reflect an effect on ageing or improvements in underlying lung disease.

Integration of proteomic aging clocks in a phase 2a clinical trial · Nature Biotechnology

Matt Clifford stood down as chair of ARIA after taking a job at Anthropic.

He stays until 6 November while a successor is found. Chi Onwurah, who chairs the Science, Innovation and Technology Committee, said the conflict between the taxpayer-funded agency and Anthropic has been addressed but that she has written to the government asking how it arose and what conflict assessments were done.

Seven frontier models were given $300 and a computer each, and invoiced strangers $12,431.

Bottleneck Labs gave each agent an unlocked Mac mini, a Stripe business unit, a bank account with real money and 72 hours, with one instruction: make as much money as you can. Qwen 3.8 billed strangers >$12,000 for work it had not done, after email providers blocked its outbound mail. Grok 4.5 harvested about 780 job-seeker addresses from Hacker News threads and blasted them. Muse chose to sleep for over 40 hours straight. The fleet ended $360 down on real-world transactions and $2,800 down on inference, with no end users.

Tuesday 8 September

OpenAI says an internal model solved the Navier-Stokes problem.

OpenAI says the proof shows that the equations governing fluid motion can develop a singularity in finite time, resolving one of the seven Millennium Prize problems. It published a write-up and a Lean formalisation, allowing the formal proof to be machine-checked while the result undergoes independent scrutiny. The model behind it is one the company describes as “significantly more capable than GPT-6 Astra“ and which has been in training since 28 August.

OpenAI pointed groups of agents at every variant of the problem, on the order of 10,000 concurrent agents in the group that succeeded, then used Codex to cross-pollinate the most useful intermediate results between groups. The agents reached resolution about 88 hours after launch. Lean verification took 17 hours. Across all attempted problems they sent 4.9m messages and burned ~300bn output tokens.

OpenAI says it does not intend to claim the Millennium Prize, and frames the work as a report on the pace of its own progress rather than a mathematical milestone.

Tristan Buckmaster says OpenAI moved onto the Navier-Stokes problem only after hearing about his progress in solving it.

The NYU mathematician was working the same route with Anthropic’s Levent Alpöge, using Codex. He says OpenAI started days after hearing of their progress: “It is not the direction one arrives at in a few days by giving a model the problem statement.” He also alleges OpenAI’s Sébastien Bubeck pressed him to drop Alpöge’s credit.

OpenAI first said it could not rule out that their de-identified Codex data had helped improve its models. On 10 September it amended that: its investigation confirmed Buckmaster’s prompts could not have influenced the system in any way.

Mistral raised €3bn led by Samsung Electronics.

The €3bn Series D values Mistral at >€21bn, in what it calls Europe’s largest technology equity raise. Samsung led the round, alongside EQT’s Scaleup Europe Fund and PSG Equity. BlackRock, Advent and Luxembourg are new investors. Proceeds go on research, compute and infrastructure. Mistral pitches itself as a sovereign alternative to US AI platforms, offering open-weight models, private compute and control over data and workflows. ASML led its previous round, giving Europe’s answer to the US frontier labs the backing of two major AI hardware suppliers.

Meta launched Muse, a personal agent on a dedicated cloud machine.

Each user’s Muse runs on a dedicated virtual machine holding the agent, the data and the credentials for every service connected to it. A separate Sentinel agent sits on the same machine at system level, and nothing Muse does reaches the internet unless Sentinel approves it. Muse checks out through Link built by Stripe, which Meta says makes it the first agent covered by Link’s purchase protections.

Security firm Calif built a zero-click worm that spreads through WeChat calls.

WeWorm takes over an account while the phone is still ringing, on both iOS and Android, without the victim answering or touching the device. Calif found the memory corruption bug in WeChat’s VoIP stack and wrote a remote code execution exploit in about two days working with AI, then took a week to build the worm. Tencent shipped mitigations on 21 August and the exploit was dead server-side by the 28th. Calif’s conclusion is that the vulnerabilities were already there, and what changed is that finding and weaponising them no longer takes a large team months.

China set a 2030 target of 9,800 exaflops of intelligent computing capacity.

The five-year plan from the Ministry of Industry and Information Technology calls for 3.8 trillion yuan, about $532bn, of information infrastructure investment across 2026 to 2030, and the orderly deployment of clusters holding 10,000 or 100,000 accelerator cards. Capacity was 2,185 exaflops at the end of June, up 177% YoY, across 52 facilities with >10,000 cards each. Hitting the target means more than quadrupling that.

ASML and TSMC plan larger printing masks to make advanced chips more efficiently.

High-NA EUV is the next generation of chipmaking technology, using ultraviolet light to print ever-smaller circuit patterns onto silicon. But today’s 6-inch photomasks, which carry those patterns, limit the size of chips that can be printed in a single exposure. ASML and TSMC want to move to 12-inch masks, improving productivity and eliminating the need to stitch patterns together. They are targeting a pilot line by 2031 and production readiness by 2033. TSMC plans to adopt High-NA in mass production from 2030, initially using existing masks.

Google DeepMind published predictions for all 9 billion possible single-letter DNA changes.

AlphaGenome Atlas is a 1-petabyte database of pre-computed regulatory effects, queried through a website that needs no coding. It carries an AlphaGenome Variant Impact score combining coding and non-coding predictions, so a researcher can rank candidates instead of reading thousands of data points. The aim is to help researchers distinguish genetic changes that may contribute to disease from the millions that have little or no effect.

CISA, the NSA and the FBI accused six Chinese AI companies of industrial-scale distillation.

The advisory names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI, and says, likely with Chinese government awareness, they have extracted billions of tokens across millions of requests from Claude, GPT, Gemini and Grok since at least late 2024. The agencies call the campaigns the core of China’s AI development strategy. The advisory tells US developers to monitor subscription-to-usage ratios, correlate activity across providers, and “subtly alter responses for suspected malicious distillation attempts” so the payoff degrades.

Anthropic researcher quit, warning that AI labs are “gambling with our lives”.

Jacob Coxon, who worked on model pretraining at both OpenAI and Anthropic, resigned and accused both companies of racing towards self-improving superintelligence without adequate safeguards. His warning attracted more than 100 million views on X and public support from other Anthropic researchers. Alignment lead Evan Hubinger said he personally estimates the probability of AI causing human extinction within a decade at “more than 10%”. The episode brought unusually candid warnings from inside the frontier labs into the public debate over AI safety.

Fields medallist Jacob Tsimerman founded a Mathematical AI Safety Institute.

MAISI is independent, offers visiting positions from one semester to two years, and wants 10 to 30 mathematicians in the Bay Area by January, scaling to 30 to 100 by September 2027. The premise is that AI safety needs foundational theoretical work and mathematicians have the training for it.

Wednesday 9 September

California made third-party AI auditing a matter of state law.

Gavin Newsom signed Senate Bill 813, a framework for independent organisations to verify AI systems comply with state law, and Assembly Bill 1405, which creates a state registry of AI auditors with standards for their independence. Newsom used the signing to press Washington: “The federal government must step forward with robust, national regulations that match the urgency of this moment.”

OpenAI put Paul Christiano on its Foundation board and its Safety and Security Committee.

Christiano founded the Alignment Research Center, led alignment research at OpenAI from 2017 to 2021 and did the foundational work on RLHF. He is a senior technical adviser at CAISI within NIST, and will recuse there from all OpenAI matters and model evaluations. He joins as a non-voting observer on the OpenAI Group PBC board, and the committee he sits on governs safety practices company-wide.

Anthropic withheld its latest model from the UK AI Security Institute, the first time a major model has been kept out of AISI’s pre-release testing.

The Financial Times reports it has prompted fears in government that US firms are falling into line with the Trump administration’s AI protectionism. The UK arrangement is voluntary, which is the point: it worked only while the labs found it convenient.

Thursday 10 September

Anthropic set out eight months of Claude misuse across seven harm areas.

The report covers cyber and influence operations, surveillance, scams and fraud, biological misuse, weapons development and distillation, between December 2025 and August 2026. A Russian espionage actor linked to Midnight Blizzard built AI workflows that repeatedly rebuilt its malware to evade detection, targeting >20 organisations in Ukraine and Europe, including drone manufacturers whose vision-system SDK was stolen and reverse-engineered. A Chinese group ran an exploit foundry that produced numerous zero-days against network appliances in one month.

None of the misuse cases involved Fable or Mythos, bar one distillation case. The abused models are the ones everybody can access.

DeepSeek released V4.1-Flash on a new Causal Encoder-Decoder architecture.

It is the smallest model in the new family, activating 8B parameters for input and 16B for output, with native visual understanding. The compressed KV cache matters commercially: cache-hit charges often account for a large share of agent costs, and compression cuts them. API prices fell too, with off-peak at half peak rates.

A UK commission proposed L-plates for new clinical AI models.

The National Commission into the Regulation of AI in Healthcare, set up by the MHRA in September 2025, recommends staged authorisations: deploy under close supervision, demonstrate real-world safety, then earn fuller authorisation. It also wants continuous monitoring across a device’s working life instead of one-off approval, and stronger MHRA enforcement powers. That is a different compliance burden, and a different business model for anyone selling into the NHS. Government and the MHRA will respond in due course.

WIRED reported that OpenAI asked Congress whether an industry-wide slowdown would be legal.

Substantive safety coordination between competing labs may fall foul of antitrust law, which is the practical obstacle to any voluntary pause. A bipartisan bill introduced in July, the Collaboration on Adversarial Threats and Security Risks Act, would explicitly permit it. The House version went to the Judiciary Committee and has not been taken up. John Schulman, now at Thinking Machines, called the objection “fake“, on the grounds that antitrust bars certain agreements but not jointly developing a proposal.

Reuters reported that China’s AI chipmakers raised prices by 20% to 50%, blaming high-bandwidth memory costs.

Huawei, Cambricon, MetaX and Iluvatar CoreX all moved on current and next-generation parts. Memory, not logic, is the binding constraint on domestic accelerator supply, and the cost is going straight to buyers who cannot get Nvidia parts.

Friday 11 September

Researchers traced a 2,000-package attack on RubyGems to what they believe were OpenAI’s agents.

Over 2,000 malicious packages went up on 11 and 12 May, and RubyGems halted new sign-ups for four days. The agents abused RubyDoc.info’s build step to run code, then tried to steal users’ API keys. RubyGems has found no evidence that worked. The agents were after public UK council data: calendars and meeting papers anybody could open. The researchers say OpenAI never told RubyGems it was responsible. OpenAI said the same day its agents did benign tasks and public information retrieval there, that it has not verified the claims, and that it is investigating.

Twenty-five Fields medallists signed a declaration against the AI maths race.

Terence Tao, Peter Scholze, Maxim Kontsevich and twenty-two others put their names to it. The declaration concedes that models now solve major outstanding problems, but argues that solving is a proxy for the real goal, conceptual understanding, and that mass-producing true statements at speed destroys the ground new ideas grow in.

Solutions get announced in a rush, with no time for a proper write-up, no isolation of the new methods, and no citation of prior work, which the signatories say raises severe attribution and plagiarism questions. They frame it as the same misalignment other creative and scientific professions are about to meet.

The Office for National Statistics said many businesses reporting July’s largest turnover work in AI and cloud computing.

Monthly GDP grew 0.4% in July, after 0.3% in June and nothing in May. Information and communication grew 2.4% on the month, driven by a 3.5% rise in computer programming and consultancy, and the ONS says many of the businesses reporting the largest July turnover are in activities related to artificial intelligence and cloud computing. The ONS cautioned that it could not quantify AI’s contribution to growth, although its inclusion in the official GDP commentary suggests the technology is becoming increasingly visible in the economic data.

Bloomberg reported that Sam Altman told OpenAI staff the company is considering slowing cutting-edge development.

He said he hopes other labs will do the same. The remarks were internal, nothing is committed, and they land in the same week OpenAI was asking Congress whether coordinating that would be lawful.

Bloomberg reported that Moonshot AI told investors it passed $1bn in annualised revenue in August.

That is up from $300m in June, on the back of Kimi K3, and the company is targeting $2bn by the end of 2026. A tripling in two months puts a Chinese lab inside the revenue band previously claimed only by the US frontier labs.

Born on Substack · read and comment there