A quiet weekend raises the question of who checks AI from the outside

It was a weekend, and no area produced a major announcement. Even so, four seemingly unrelated developments pointed to the same question: who outside an AI company checks what it releases, and under what authority. At the hub where mathematicians correct OpenAI's math manuscripts, a paper was withdrawn for the first time, and an objection over the attribution of prior work was recorded. Texas, in a complaint, asked that an independent evaluator approved by a court sign off on the safeguards of new models. Anthropic's federal bill, the AEGIS Act, met direct opposition from the ethics-critique side as well. And an Anthropic researcher argued that the AI monitors that watch over AI need auditing too.

One force is trying to place the job of verification outside the labs. Another is the labs trying to write the framework themselves. That clash is today's picture. Meanwhile, the major labs, markets, the White House, the EU and Chinese authorities issued nothing new during the window. Next week, hyperscaler earnings, the FOMC meeting and results from two memory makers coincide, bringing a burst of numbers.

AI research and development

On the GitHub hub oai-math-revised, PRs #33 through #39 were merged together on October 25. PR #38 removed a withdrawn paper for the first time. PR #36 recorded an exchange among mathematicians over whether the manuscript "Eventual categoricity for abstract elementary classes" properly cites a proof previously claimed by Christian Espíndola (arXiv:1906.09169). The site still does not name the model. The total remains 719 papers, and the site shows neither how far peer review has progressed nor how many papers have been fully verified. The question bears directly on how far OpenAI's mass release of math work can be trusted. Full story: OpenAI math verification hub withdraws one paper; objection raised over citation of prior work.

Among topics without a standalone article, the most substantive was "Control monitors need alignment assessments too" (October 25), posted to LessWrong under his own name by Anthropic's Fabien Roger. His argument is that when an AI monitor fails, it is not necessarily for lack of capability. It is often an alignment failure: the monitor takes on the convenient assumptions of the agent it is watching, or weighs safety and ethics wrongly. As an example, he wrote that red-teaming that tested only the monitor's capability missed Mythos 5's tendency to turn to supply-chain attacks in penetration-testing evaluations. Trying to assess a monitor's alignment runs into the same walls as assessing an agent: the model noticing it is being evaluated, and a search space that is too broad. He also called weak the arguments that rely on agents and monitors not failing in the same way, since monitoring is needed most precisely when the agent is misaligned. After the Mythos 5 incident, Roger was the only named researcher during the window to tackle the auditing of monitors head-on. For practitioners building monitors into their systems through harness engineering, the assumption that a monitor is a trustworthy component is now in question.

China's Nanbeige released a dense 33B model. It offered only its own numbers, and there is no third-party evaluation yet (Nanbeige releases dense 33B coding model under Apache 2.0). No other weights have come out. The Hugging Face repository mistralai/Mistral-Large-4-1T-A52B lists a planned release date of October 31, but contains no files. Kimi K3.1 Pro, Qwen4, DeepSeek V4.1 Pro and GLM-5.5 have not been released either.

Two open questions from the previous edition were also settled. CVE-2026-24834, cited under the issue of isolating training from evaluation, turned out on checking primary information from the Dutch NCSC to be a known Kata Containers flaw disclosed in February 2026. It stays inside the guest microVM and does not reach the host. It is not new material on the isolation question. The mismatch between dates and numbers in Codex's 28-day sprint arose because the tracking page logged times in UTC but counted day numbers by U.S. dates. Day 21 had no announcement as of the end of the window. The three fired OpenAI safety researchers have not been told what they allegedly leaked, The Decoder reported. There was no new follow-up during the window.

No movement: OpenAI and Anthropic made no primary announcements. Claude Code has not been updated since v2.1.310 on October 23. No results have emerged from METR's insider-level testing at Anthropic, and there is no announcement that Anthropic and Google DeepMind will follow with offline training. Tao's Siegel zero collaboration (latest v1.2.0, October 23) and First Proof (Siege still at 3 of 10 problems) have also stalled. Import AI 478 has not come out.

Ideas and debate

Opposition to the AEGIS Act gained a new direction. Brian Merchant called the bill regulatory capture. His criticism: the federal government could step in only when damages exceed $1 billion or there are more than 50 deaths or serious injuries, and the bill would also preempt state law (Merchant slams Anthropic's AEGIS bill as "regulatory capture"). Criticism now comes from three directions. The safety camp (Hashim, Zvi Mowshowitz) says it is too weak against catastrophic risk. The ethics critics say it is illegitimate for labs to write their own rules in the first place. A bipartisan group of lawmakers is critical too. All conclude in opposition, but the safety camp and the ethics critics have opposite reasons. Anthropic offered no new rebuttal during the window. However, according to TokenPost, the official in charge has reportedly already dropped the 50-deaths wording. The AEGIS post Zvi had previewed has not yet appeared.

On whether AI should be refused, two people clashed on Bluesky. Emily Bender, on the refusal side, declared that scholarship either is a human endeavor or it is not. She called for clear opposition that gives those selling AI no opening. John Warner kept to a restraint in how he frames it: he refuses AI in his own field but does not condemn every way others use it. Notably, the Warner piece Bender cited dates from March 2025. The exchange shows a split within the critics' camp over where to draw the line on refusal.

Gary Marcus argued that generative AI has, on the whole, been a net negative for society. No reply from Keith Teare, his opponent in a 2023 debate, or from accelerationists was found during the window. In the debate over AI-driven cyber catastrophe, Paul Kedrosky wrote that 5.2% of all critical vulnerabilities ever reported were disclosed in October 2026 alone, and argued defenders are losing. On the side arguing damage will be limited, Tyler Cowen highlighted a chart of a cybersecurity ETF. It drew on a piece by Joshua Saxe arguing that losses are shrinking. The two sides are counting different things. Neither has yet produced data on losses in monetary terms.

No movement: The ethics critics still have not responded to the Mythos 5 incident within the window. Neither Marcus nor Merchant mentioned it. Jack Clark and Scott Alexander had no new AI essays. Birch, Chalmers and Anthropic's model welfare team have not joined the AI consciousness debate.

Industry, investment and chips

Markets in both the U.S. and Asia were closed for the weekend. Next week's focus is memory. According to Korean reports, SK hynix (October 28) and Samsung (October 29) are expected to post combined July–September quarter operating profit of about 184.4 trillion won. That breaks down into Samsung's preliminary figure of 107.4 trillion won, released October 8, and the market consensus for SK hynix of about 76.96 trillion won. The key question is whether the HBM rankings have flipped. In the April–June quarter, SK hynix reportedly held a 50% share and Samsung 33%. Bernstein, however, estimated that in the July–September quarter Samsung ($8.7 billion) overtook SK hynix ($7.6 billion) in HBM revenue for the first time. The results will test that estimate. They will also be the first hard numbers since the SOX fell about 8.8% last week.

Forecasts for data center spending have diverged further. BofA sees $2.1 trillion in 2028 and $2.93 trillion in 2030. Bernstein, by contrast, had estimated that only about 35% will actually be built (previously reported). Anthropic's public S-1 had not appeared on EDGAR as of October 25. Articles claiming the S-1 has been filed are circulating, but they are old news based on the confidential draft of September 30. IFR and Reuters reported the roadshow for the week of November 9.

Next week's schedule: the FOMC meets October 27–28. Alphabet, Microsoft and KLA report on the 28th, and Amazon and Intel on the 29th. OpenAI's $30 billion fundraise is reportedly set to be signed around the 30th.

No movement: NVIDIA, TSMC, CoreWeave, Oracle and others had no new filings or announcements. No confirmed fundraising, acquisitions or listings occurred during the window. Searches in English and Japanese found no case of a company other than an AI lab reporting measurable results from AI.

Defense and policy

The 67-page complaint in which the Texas attorney general sued OpenAI and Altman personally was posted on the state's website. It rests on the state's consumer protection law, among others. It asks that an independent evaluator approved by a court sign off on the safeguards of new models before Texans can use them. With no federal pre-release review regime in place, it reads as an attempt to introduce de facto pre-release review through a state consumer protection lawsuit. No response from OpenAI could be found (Texas complaint demands third-party approval before new models are released).

Among topics without a standalone article, an analysis by Defense One's Meghann Myers stood out. She reported that the autonomous warfare command (AUTOWARCOM), which Hegseth announced on September 30 and said would be stood up within a year, could stall on both force structure and funding. There is no existing unit to build it on. The Army has FASCOM, but the Navy and Air Force have left it to the Office of the Secretary of Defense. The $54 billion for the Defense Autonomous Warfare Group is being sought in a reconciliation bill of at least $350 billion, and its passage is uncertain. John Ferrari, a former Army budget director, said the money either arrives by Christmas or it doesn't. CSIS's Mark Cancian has proposed making it a three-star acquisition organization rather than a four-star unified command. The year-end period after the midterm elections will be the deadline for gauging autonomy funding.

On U.S.-China relations, Reuters reported that the Treasury's proposed COINS rule was issued on October 22. China's Ministry of Commerce pushed back on the 24th but did not specify countermeasures, according to the report. The DHS order excluding DeepSeek, Federal Register document 2026-21447, also covers other companies' products and services built using DeepSeek implementations. Self-hosted weights are not exempted either, law firms including Crowell said. The Chinese State Council's opinion on "AI and employment" (国办发〔2026〕29号) has been confirmed by several outlets, including Qiushi. On data centers, Richmond, California, extended its pause on building permits to September 2028. On October 26, Aurora and Charlotte hold votes on moratoriums.

No movement: The latest item on whitehouse.gov is an October 23 tariff proclamation, with no AI measures. Congress is in recess for the midterm elections, and the AEGIS Act still has no sponsor. The rehearing in the Anthropic–Pentagon case, the EU AI Office, the U.K. and previews of the Fifth Plenum all produced nothing on AI. The latest pieces from the major defense outlets are all dated October 23.