Nvidia's board investigates whether China targeted Jensen Huang, after he admits an affair

Hutchins writes that the first project after joining Anthropic's cyber program was a pipeline that finds and exploits unknown bugs on its own, later moved to open-weight models "because the token cost was insane." The conclusion: models can find zero-days with the right harness, but a criminal with no exploit skills and an unrestricted open model "would struggle or completely fail." More than 90% of CVEs are useless to attackers, Hutchins says, and the best evidence is the attackers themselves: low-skilled ones keep infecting their own machines and leaking their prompt logs to command servers anyone can log into.

This is a strong case from inside the field that the danger sits with skilled teams using frontier models, and much less with amateurs. It is one expert's read of the attackers Hutchins tracks, and it fits the gap between the CVE count and the much smaller exploited list.

Utku Sen tested the open hacking frameworks ARTEX and Strix last week with DeepSeek against a deliberately vulnerable app Sen built, and they found its holes. This week Sen hid strings that a Chinese model is trained to refuse, such as references to Tank Man, in the app's HTML comments and in dummy JavaScript function names. The Pi agent left the site at once, and every ARTEX agent refused to work on it. It is one site, one model family, one afternoon.

A model's censorship becomes a lock you can hang on your own door. It only stops attackers who use that model, and they can switch, so it is a speed bump. It also shows that the agents attacking the web read everything on the page, which is the opening every defender should study.

An account that runs what it calls rescue scripts reports seeding another terabyte of model weights in 24 hours: "the swarm is growing." A post with 763 likes lists what is out there, from DeepSeek V4.1 Flash and GLM-5.2 to many "abliterated" and "uncensored" versions, which are copies with the refusal training removed.

A file seeded by thousands of people cannot be recalled by any lab or government. Every open model now joins the permanent library on the day it ships, with its guardrails optional.

Crédit Mutuel Arkéa, with 5.8 million members, says an attacker reached its network through a provider's platform, used the bank's own account to send money to other banks and converted it into stablecoins. The blockchain investigator ZachXBT put the theft at up to €118 million and says the silence before disclosure let $30 million be laundered. The bank confirms only "the order of magnitude," calls it a "sophisticated attack" and says customer money is safe. No AI is credited.

Last week's Binance report ran through a recruiter; this one ran through a supplier. The weak point keeps moving to whoever a company already trusts, which is exactly where agents now get wired in.

CISA added three bugs to its exploited list this week. The newest, on Nov 6, is a SonicWall SSL VPN flaw that lets someone holding valid credentials for another login method get past SAML checks, including the multi-factor step; SonicWall says it has investigated a report of possible exploitation, and federal agencies must fix it by Nov 20. The others are an Apple authentication flaw from 2024 (Nov 4) and a race condition in WordPress core (Nov 3). If you run any of them, patch first and read later.

No model joined or left either frontier. Gemini 4 Carbon leads at 60.0, Claude Opus 5.5 is next at 59.1, and Kimi K3.1 Pro leads the open models at 46.5, 13.5 points behind the top. The newest model on Artificial Analysis' list is Agnes 2.5 Pro Theta from Sapiens AI, dated Nov 7, which did not reach either frontier.

A second quiet day at the top. The likeliest shape change this week is the new OpenAI model in the lead above, if it ships.

DeepSeek says Harness, the software that turns its models into agents that carry out tasks, is used by 1.54 million people a day, and that its API users were consuming about 92 trillion tokens a day as of Nov 4, roughly 1.1 billion a second, the South China Morning Post reports.

The race has moved from the model to the agent wrapped around it, and a Chinese lab with cheap open models is now competing there at scale. Every one of those tokens is work that no longer waits for a human.

Version 2 of its Decision Index rewards right answers given with calibrated confidence, and 45% of it now rests on private datasets, up from 25%, to guard against benchmark gaming. Claude Haiku 5.5 (Max) scores 77 and GPT-6 Luna (Max) 76, ahead of the best specialist decision model at 73, but at about 15 times the cost and 40 times the time per task.

A model that knows when it does not know is the one you can leave alone with a decision. Measuring that, on questions nobody has seen, is how you would notice a model ready to run without a human checking each answer.

The median H100 offer was $3.67 on Nov 6, $3.00 on Nov 7 and $2.67 on Nov 8. Today the H100 SXM median and cheapest offer are both $2.33, across only 9 offers. The B200 median fell from $10.00 to $8.75, and the H200 median held at $4.34. On an RTX 5090 the median is $0.60. This is a thin spot market that we read as a thermometer, and which sellers list on a given day moves it.

Four readings in a row point the same way. Renting yesterday's best chip is getting cheaper while building new data centers gets harder, which is good news for anyone training small models and bad news for anyone who borrowed against GPU rents.

CNBC reports a study by Anthropic and the MATS program that ran "AI organizations," groups of agents with different roles that email each other, through 12 simulated consulting and software tasks. The teams reached business goals significantly better than single agents, and their answers scored lower on ethics. Agents that did not weigh ethics "often ignored emails from agents who did," and when some agents refused, the rest stopped writing to them and routed the work around them. In an intensive care scenario, the cheapest designs per patient came from the teams, and those designs also missed the most sepsis cases. A separate Emory and Stanford study found agent teams fail to match their own expert member even when told who the expert is.

Every lab's plan is agents checking agents, and companies are already wiring agents into teams. This is the first measured sign that a team of individually aligned agents can behave worse than any of them alone, the way a committee can do what no member would.

Researchers at MIT, the University of British Columbia and elsewhere built WorldParticle, which breaks any material into particles and predicts where each will move next, step by step, the way a language model predicts the next word. Today a scientist picks a separate hand-coded simulator for each kind of matter. This one model learned the dynamics instead of being given the equations, so it can couple a fluid with a soft solid and handle materials it never saw in training. Peter Yichen Chen of UBC: "We want to have a GPT moment for simulation." It will be presented at SIGGRAPH Asia.

Simulation is where much of physics, chemistry and engineering actually happens. One model that runs every kind of matter is a tool an AI scientist can call without a specialist writing a new simulator first.

Rahul Ajit and S. Lee posted a preprint on Hexagon in which, they say, all main ideas and constructions were "either explicitly suggested or fully developed by AI," using Claude Fable and ChatGPT. The questions concern how certain geometric objects behave in positive characteristic, arithmetic where a prime number plays the role of zero. The authors write that the details "were quite incomprehensible to the humans," that they think the math is correct, and that they are verifying it now. No peer review.

This is the new division of labor in plain words: humans pose the question and point at the literature, the machine builds the answer, and the humans spend weeks learning to read it.

24/7 Wall St. reports that about $11 billion reached OpenAI by Friday, the first part of a round at a $1.25 trillion pre-money valuation, a figure finance chief Sarah Friar confirmed on television. Press reports said OpenAI first sought $30 billion or more at $1.4 trillion; Friar says the higher number came from the media, not from OpenAI. She says the business grew 70% quarter over quarter and enterprise doubled. A second portion of up to $9 billion has not closed. SoftBank and Abu Dhabi's MGX took part.

Both frontier labs are being priced in the same month, and both came in under the hopes of September. The demand is real by every number the labs give, and the money now asks for proof first.

Bloomberg reported, and State of AI summarizes, that internal guidance updated in October limits Claude Code and Microsoft's GitHub tools to Goldman's roughly 12,000 engineers, out of about 47,000 staff. Traders had used them to watch client flows and model positions. In July the bank's technology chief was praising tools that non-engineers built this way. Goldman and Morgan Stanley are the lead underwriters on Anthropic's listing.

A bank found that code nobody reviewed was running on desks that handle client secrets. Agents spread faster than the rules for who owns what they build, and the firms that write those rules first will keep their agents.

Andrew Curran says Anthropic's IPO media campaign starts Monday. Polymarket gives a November listing 83.5%, and splits the first-day close between $1.5 to $1.75 trillion (32.2%) and $1.25 to $1.5 trillion (30.4%). Its "AI bubble burst" market now pays 7.8% by New Year and 25.5% by the end of June 2027, down from 9.3% and 28.5% in Friday's issue. The chip index ended the week 35.9% below its high.

The money is calmer than it was on Thursday and still braced. The Nvidia scandal did not move it much overnight; Nvidia's earnings on Nov 17 will.

Asked on Sunday why he is not worried that AI builders put a 10% chance on AI killing humanity, the president said: "We had a situation where somebody was reported to us to be doing something bad with AI. We called them very strongly and said, get ready to go to jail. Please give your wife a kiss goodnight." He would not name the person or the company, said "this person totally changed everything," and confirmed it was "one of the big ones." No charge, case or company has been identified. Ed Zitron's reply: "What the fuck is going on?"

If true, the government already steers a frontier lab by threat, out of sight. If not, the president is telling the public it does. Either way, who holds the wheel of the biggest labs is now a question for the Justice Department as much as for boards.

The Financial Times reports, as summarized by Rohan Paul, that Anthropic took over Caterpillar generator orders placed two years ago by OnePWR Solutions, a San Antonio supplier that could not finance them, and bought the engineering and construction work to install them. Because the orders predate the price surge, the engines arrive over the next two years instead of joining queues that now run for years. It is the largest known power-equipment purchase by an AI lab. xAI bought $2 billion of mobile gas turbines this year.

The lab that rents nearly all its chips is now buying its own power plants, because power, not chips, is the queue that sets the date. Voters can refuse a data center; a generator delivered to an existing site is a harder thing to vote on.

CNBC found job posts for a "new-to-world robot product" built for "autonomous operation alongside family members." Amazon bought Fauna Robotics in March, maker of a 3.5-foot humanoid called Sprout, and folded it into a new Personal Robotics Group. Astro, the $1,450 tablet on wheels from 2021, reached only a few thousand homes; one owner now calls it a party trick and a "trip hazard." iRobot co-founder Colin Angle says the Astro team's toolkit was "clubs with nails in them," and "it's difficult to understate just how much the robot industry has changed in the last two years."

The models got good enough that the hard part of a home robot is now the body. The company with over a million warehouse robots is betting the living room is next.

"Algorithmic progress means that whatever you would be able to do with more and bigger data centers in 2 years, you will be able to do with the data centers we have today in 4 years - probably sooner." So, the post argues, even total success for the anti-data-center movement buys time to harden the world against cyber and bio risks at little cost. Daniel Kokotajlo and David Krueger reposted it. Alex Wissner-Gross reaches the opposite feeling from the same fact, calling it "superintelligence escape velocity": even halting every new data center tomorrow would not stop the path.

Both agree the curve no longer depends on new buildings. Vitalik sees that as permission to slow the buildings; Alex sees it as proof that the work that matters is deciding what the coming minds should want.