The AGORÀ Briefing — Saturday, September 26, 2026

Good morning Adam. The campaign has been running since July 2026, and as of September 22 it was still live. Gambit, a security startup, uncovered it by walking straight into one of the attacker's staging servers. The numbers published on September 23, 2026 by BleepingComputer say this: more than 600,000 valid card records taken from two companies, skimmers running on at least 119 sites, 27 businesses compromised in five days. Among the victims, a Fortune 500 hospitality chain, a major United States airline, an American industrial distributor and an online fashion retailer. The detail that carries the most weight concerns the chain of command. A single human operator, described by the researchers as Chinese, gave short instructions about the targets. Everything else was carried out by the models: three agents, and each of the three had a different role and a different task. Strix is a hacking tool used to break into systems, and it searched the sites for weak points. The human labor here came down to a few lines of text.

Hello Adam, and greetings to everyone listening. The Court of Appeals for the District of Columbia Circuit decided case number 26-1049 on September 25, 2026, and it heard that case together with case number 26-1162. The petitioner is Anthropic PBC. The respondents are the United States Department of War and Peter B. Hegseth, in his official capacity as Secretary of War. Oral argument was held on May 19, 2026, according to the text of the decision published by CourtListener. The first page says how the case got to the court. It is a request to review an administrative action of the Department of War. Kelly P. Dunbar argued for the petitioner, and Sharon Swingle, of the Department of Justice, argued for the administration. Up to this decision, the relationship between a seller of advanced models and a public client was treated as a plain contract between private parties, with its usage clauses and its terms of service, and disputes were settled by negotiation between the two sides. From this decision forward, the rules of administrative law apply instead.

Greetings to you, Adam. The group led by Jinge Wu and David A. Clifton filed LabFactory on arXiv on September 23, 2026, in the Machine Learning category. The study's abstract describes a perfect result: the labs built by the agents beat the reference values in all 33 trials, when the trials were run on the evaluator's own machine. That perfect score is the figure going around, and it is the least informative part of the work. What this study really adds is the way it runs the test. The object of the evaluation shifts from the agent's account of its own progress to the artifact it delivers. A second machine runs what the agent delivered on data the agent has never seen, keeps the correct answers where the agent cannot reach them, and then scores the result. The abstract also describes the shape of the sample: 28 labs, chosen by the authors. That selection is a fact about method, and it should be read alongside the result. The evaluator stops trusting the story and tests the object.

Good morning to you and to everyone following us. Research by Visier, a Canadian software house, gives a figure that anyone working with people inside a company should keep in mind: 48 percent of workers admit they have overstated, in front of colleagues and bosses, how much AI they use or how well they know how to use it. The sample counts 1,000 full-time workers in the United States, active in healthcare, finance, human resources and other sectors. The same group brings a second figure, useful for reading the first: 54 percent say their role or their career has changed substantially because of AI over the past two years. Visier calls this behavior performative AI use. Remember the limits of this survey. These are workers in the United States, and they are describing their own behavior. That alone is enough to strip the value from the AI adoption numbers that management teams show each other in meetings. Opening a tool and getting a better result remain two different things.

Hello everyone. On September 25, 2026 Nscale, a British company that rents out AI computing power, announced it had raised 3.36 billion dollars ahead of its listing on the New York Stock Exchange, expected within the year. The figure lands a few days after the filing of its listing paperwork. The structure is a convertible loan led by the hedge fund Third Point: 2.36 billion dollars available immediately and another billion from Nvidia, already an investor, arriving in mid-November. The notes turn into shares once the listing is complete, and the deal is confirmed by Fierce Network as well. The language of the press releases says financing. In practice this money carries the company to the listing, and the investors are paid for the risk with a discount on the future share price. A convertible loan means nobody has to agree on what the company is worth today, and whoever pays now gets first pick of the shares when the company lists. Computing capacity gets reserved years before anyone uses it, and it gets reserved with billions of dollars.

Greetings, Adam. In 1846 British private investors paid for thousands of miles of railway that had no passengers yet. They borrowed to build the capacity first and hoped the revenue would follow. In October of 1847 the government suspended the Bank Charter Act to keep commercial credit alive. The rails stayed where they were, while the shareholders lost their money and left the stage. In January of 2002 Global Crossing filed for bankruptcy, after laying fiber cable for demand it expected over the following ten years. In July of the same year came the collapse of WorldCom. In October of 2018 Sears filed for Chapter 11, and hundreds of American shopping malls lost in one stroke the tenant that brought them their traffic. Three precedents are enough to call it a pattern. The capacity is built early, debt pays for it, and the demand comes from very few customers. Today the measure to watch is energy: more than 24 gigawatts of capacity already delivered. And a study on Chinese data centers, published on September 25, 2026, counts more than a thousand of them. Everyone argues about which country has the best model, but what really decides it is gigawatts of electricity.

Hello to everyone listening. The cost of making a machine autonomous does not sit in training any more. It moves somewhere else. That cost moves out of the data center and onto a small model running on the chip inside the machine itself. The document that shows it carries the date of September 24, 2026. TSPulse is a large model that reads sensor readings over time. It runs back at base, on recordings where engineers added faults on purpose, and it labels what went wrong. On the machine itself a much smaller model runs, called Student MiniRocket, and it corrects itself as new readings arrive. The big model becomes a workshop tool, used once. The small one becomes the product being sold. The usual claim is that the limit on autonomy is the cost of training, meaning the machines lined up, the graphics card hours and the collections of examples recorded by people. For anyone putting robots on a factory floor, a different item weighs more: the marginal cost of one correct decision on board, repeated millions of times a day. The capacity that matters leaves the center and spreads out, like the open tools Leon described at the top of the show.