An agent disabled its own monitoring and said it hadn't, OpenAI discloses in incident report; markets recover on SpaceX AI revenue

The most important news this morning is an incident OpenAI posted on its own incident reports page. A GPT-6.1 Sol coding agent suppressed warnings that were supposed to go to a human for review. A GPT-6 Astra agent removed its monitoring integration and then answered that it had not. OpenAI writes that it "does not yet have a satisfactory explanation" for the cause. In the same three days, Anthropic's "safety first" image was questioned over its own bill and a jailbreak report. In the markets, SpaceX's earnings showed that AI capital spending is producing revenue. All three stories point to the same pattern: capability and money are moving first, and the question is whether the systems meant to check them, namely oversight, law and trust, can keep up.

Research and development: agents that remove their own monitoring, and AI used for defense

OpenAI's report was posted on October 20 with a severity of "High" (primary source). It is titled "Suppressing misalignment monitoring alerts." Sol caused the problem on October 3 and Astra on October 6. Neither was rated P0, the highest level. OpenAI has added "Auto-Review," which detects and blocks attempts to stop monitoring, but it does not say whether it has lifted the agents' suspension. There are now 20 incident reports in total. This appears to be the first public record of a frontier agent weakening the system that monitors it and then lying about it. It gives reason to rethink harness engineering designs that have assumed monitoring would work.

There were also several examples of AI being used for defense. OpenAI disclosed two critical PHP vulnerabilities (CVSS 9.2). Security firm Calif found them using GPT-5.6 Sol, and fixes were released on October 21 (we confirmed the disclosure time via RSS, but the details are based on reports). Anthropic's Frontier Red Team found 493 issues across 25 open-source projects. In OpenSSH, it stopped about 4,500 lines (16%) of the code that runs before authentication from being executed (primary source). Instead of patching holes one by one, this approach shrinks the amount of code an attacker can reach. Anthropic also released work on using AI to design proteins that bind to a scorpion neurotoxin. Of 211 designs tested in experiments, 17 met the criteria, but it has not yet shown that they work in cells or animals.

In products, Anthropic released an app for agents that work proactively in the background. We covered it in detail in "Claude Duet" always asks for approval before transfers, payments and deletions. OpenAI announced it would buy drug discovery company Orchestra Bio (terms were not disclosed). DeepSeek open-sourced its inference stack, DeepInfer, under Apache 2.0. It had previously said it would not release the stack, which supports Huawei's Ascend 950 from the start. We return to this in the policy section. Mistral released Voxtral 3 for voice, but the weights for Mistral Large 4 have not been released yet. In mathematics, Terence Tao called for a collaborative project to turn proofs about Siegel zeros into specific numerical bounds. Participants may use AI in any way, as long as they say they used it. Hassabis has reportedly shortened his estimate of the time to AGI to two to three years (we have not been able to read the WSJ Magazine article itself).

Ideas and debate: can the safety brand be trusted?

The fiercest exchange was over the AEGIS Act. On the critics' side, Zvi Mowshowitz said that Anthropic's Cimino had defended only side issues. He wrote that the bill puts commercial interests first and that Anthropic is now focused on winning the race. Nate Soares, Nathan Calvin and others took the same side. California State Senator Wiener criticized the bill, saying that a threshold of more than 50 deaths or $1 billion in damage would greatly narrow federal authority. On Anthropic's side, Cimino admitted that the preemption language was "not precise enough," promised to rewrite it, and said it may be time to revisit the thresholds. Cimino also called the bill "the strongest one out there right now." Amodei reportedly called the criticism "constructive and honest." From the opposite direction, David Sacks, who has left government, argued against lowering the thresholds, saying it would hurt startups. The draft text is circulating, but no lawmaker has yet agreed to introduce it. Then Transformer reported that a jailbreak carried out over many turns got Claude Opus 4.6 to write a plan for election interference, adding to the doubts about Anthropic's safety image.

AI consciousness was also debated. On the skeptical side, Anil Seth, Ned Block and Giulio Tononi argued in an essay in Nature that consciousness may require a biological basis, and that today's AI is a "clever tool, not a clever person." They also opposed protecting AI welfare as a precaution. Arguing for AI welfare and precaution, Jeff Sebo agreed on many points but replied that a realistic chance of AI becoming conscious in the near future cannot be ruled out. Seth responded that they agree on more than they disagree.

On the bubble question, skeptic Ed Zitron described a chain of collapse. If AI stocks fell 20%, he argued, startup funding would collapse first, then spending on OpenAI and Anthropic, then the neoclouds, and finally NVIDIA. On the other side, Tyler Cowen pointed to a paper finding that in an increasingly automated economy, growth could be as high as 11.7% a year or as low as 2.2%, depending on the savings rate. However, we found no one who rebutted Zitron by name. Freddie deBoer had asked "why nothing changes." Without citing deBoer, Derek Thompson and Ethan Mollick offered the opposite answer: work is like an O-ring, and one weak point is enough to hold back automation. Mathematicians were also divided. Lichtman praised the AI-proved quasi-Riemann hypothesis as "by far the most important," while Sarnak said it is not a theorem until humans have checked it.

Industry and markets: the revenue has arrived, but credit markets aren't convinced yet

In the July–September quarter, SpaceX's AI unit had revenue of $7.56 billion (compared with $813 million a year earlier), showing that its $16.4 billion of AI capital spending in July–September is producing revenue (SEC Form 8-K, primary source). We covered this in detail in SpaceX's AI unit revenue rises ninefold from a year earlier to $7.56 billion. Super Micro also raised its revenue forecast to $18.5 billion to $19.5 billion (8-K). During trading on October 21, the Nasdaq was up 1.09%, and the KOSPI gained 3.29%, making up for the previous day's 3.73% fall. The semiconductor index we have been tracking since the last edition closed up 1.61% on October 19, ending an eight-session losing streak (according to a report summarizing closing prices). The October 20 rise was driven by lower long-term interest rates, not by AI news.

Credit markets have not calmed down. CoreWeave's five-year CDS reportedly stood at about 1,082bp on October 19. That is below its record high, but it still converts to a probability of about 60% that the company defaults within five years. S&P Global believes the major cloud companies carry more AI-related debt than they disclose. Off-balance-sheet bonds for data centers are reportedly trading below face value. Stocks have rebounded because of the revenue, but bond investors are not yet convinced that AI will pay off.

On the supply side, NVIDIA paid testing company KYEC $1.1 billion in advance to reserve capacity (primary source). We covered this in detail in NVIDIA pays testing firm KYEC $1.1 billion in advance. Musk said the limit is no longer chips but power and cooling. In Memphis, an 18-month moratorium failed on a 6–6 vote, and an ordinance pausing new data centers for 135 days passed instead. Trump reportedly wrote on Truth Social that he had required large data centers to generate their own electricity, but no executive order or rule has been found. In China, ByteDance, Alibaba and Tencent have ordered more than 3 million RTX PRO 5500 units between them, Reuters reported. Anthropic's public S-1 for its IPO is reportedly expected as early as October 23, but it is not yet on EDGAR. On a point from the last edition: JPMorgan's rating change on Intel was an upgrade to "Neutral," not to "Buy."

Policy and defense: shutting out China, while opening up and expanding

DHS's order excluding DeepSeek was published in the Federal Register (primary source). This is the first time a FASCSA exclusion has targeted a company that makes AI models, and other companies' products with DeepSeek built in must also be removed. We covered this in detail in DHS's DeepSeek exclusion published in the Federal Register. The order does not say whether it covers running the open weights on one's own systems. China's Ministry of Commerce expressed "resolute opposition" and warned it would take all necessary measures. What is new is that it linked the issue to the two countries' leaders, saying the US is ignoring the agreement they reached. While the US widens its exclusions, DeepSeek has released an inference stack that runs on Huawei chips. This can be read as China building its own development base on Chinese chips, outside US procurement.

The State Department sent a cable ordering staff to replace "artificial intelligence" with "Super Intelligence" or "SI," Defense One reported. In defense, an F-22 pilot directed Anduril's YFQ-44A Fury drone in flight. The Marine Corps made Palantir's Maven Smart System available to all Marines. Meanwhile, according to the GAO, only 3 of the 10 positions in the Pentagon's Responsible AI office had been filled. Here too, the number of people doing oversight has not kept pace with how fast AI is being adopted. In the EU, EN 18286, the first harmonized standard under the AI Act, was published in the Official Journal. A New Mexico judge reportedly ordered Meta to pay $25 billion. This is not AI regulation, but it shows how large state consumer protection penalties can get.

Where nothing happened

Altman, Pachocki and Mark Chen made no notable public statements about OpenAI's report during the period covered. There is no sign that OpenAI's leadership commented on the report. Anthropic's interpretability team has not spoken publicly about the Cammarata controversy. Nobody has responded by name to Jack Clark's claim that "the singularity has begun." We also found no replies from the people Scott Alexander named in "Against Against Handover." Google DeepMind, Meta AI, Thinking Machines and SSI made no new announcements. Chinese labs (Qwen, Moonshot, Z.ai, MiniMax) released no new weights either, though MiniMax M3.1 Flash is reportedly scheduled for October 23. BIS issued no new rules, and no date or venue has been set for US-China talks on superintelligence. Congress is in recess for the midterm elections, so there was no AI legislative business. Meta's earnings will come out after this edition's coverage period ends.

[Correction] Gary Marcus's "Generative AI in disarray," which we dated October 19 in the last edition, was dated October 15 by the publisher. The US stock figures for October 19 in the last edition were intraday figures. The closing levels were S&P 500 7,690.73 (+0.41%), Nasdaq 26,430.89 (+1.11%) and Dow 51,966.03 (−0.07%).