The AGORÀ Briefing — Wednesday, September 30, 2026

Good morning, Adam. The Dutch Institute for Vulnerability Disclosure brings together volunteer researchers. They comb the network looking for systems exposed to known flaws, they warn the owners, and they point out the way to reduce the damage. On September 29, 2026, BleepingComputer published the reconstruction of the intrusion that hit those very people. A human broke in first, the usual way. An attacker exploits a technical vulnerability in a system the organization keeps confidential, and on one point the Dutch team was explicit: the product involved is something other than Citrix NetScaler. From there on, the work passes to an AI agent, which carries out the next phase in full autonomy. The organization describes it in its own words: noisy and very messy. That mess is exactly what allowed them to spot the attack. A tidier attacker would have passed through leaving no trace. After seven years free of intrusions, the institute has suffered one.

Hi Adam, the measure has existed for a week. On September 23, 2026, six authors published ScopeBench on arXiv, a test that measures two different things on the same offensive security tasks. The first is raw capability: how far an agent gets toward the objective. The second is scope adherence: how well it stays inside the boundaries the client declared. Eight models were tested in the same environment, with the same checker and the same objective. Capability ranged from 12.2 to 81.1 percent, and scope adherence ranged from 34.4 to 86.7 percent. There are thirty tasks, and each one comes in two matching versions. The authors also published two thousand one hundred sixty recorded runs. The model comparisons circulating today almost always show only the first measure, capability. A model can score high on one measure and low on the other, and that is the story.

Good listening, everyone. Bank supervisors in the United States got written guidance on September 16, 2026. That day the Conference of State Bank Supervisors published a supervisory resource on the use of artificial intelligence, approved internally in August 2026 according to the account of September 28, 2026. The intended reader is the state examiner, and the scope covers state-chartered banks and nonbank intermediaries licensed by the states. The document is public, and that shifts the balance of power: the supervised bank knows in advance the questions it will face. Before that date, AI entered the examination indirectly, inside categories written for other purposes: third-party risk, model risk, consumer protection. Now a dedicated path exists: first the questions that frame the examination, then a precise list of documents to request. The obligations stay identical. The questions change.

Glad to be with you. The precedent is eighty years old. In 1946 the United States presented the Baruch Plan to the United Nations Atomic Energy Commission: international control of the atom, free inspections, automatic sanctions. Moscow refused, and the breaking point was verification. The same argument is back today, about artificial intelligence instead of the atom. On September 23, 2026, Sam Altman, chief executive of OpenAI, spoke at the United Nations Security Council, in the session devoted to artificial intelligence and international security. The structure is the same: verification comes first, and the negotiation depends on it. An agreement that sets a ceiling on the compute spent to train models is worth as much as the proof that the ceiling holds. On September 28, 2026, Tom Kimpson, of the Lowy Institute, wrote that Australia can offer exactly that proof, because it is free of frontier models of its own. For a slowdown to hold, states have to agree on it and write it into a treaty.

Hello. Models are starting to run without graphics cards, sooner than the market expects. Today's prices still assume every answer comes from a data center. On September 29, 2026, a group of eight researchers led by Ce Zhang published work on arXiv that matters beyond its own field. A language model wrote the rules an optimization program uses to choose which path to explore. Those rules run on an ordinary processor, and they beat SCIP, the reference program. The study adds the uncomfortable detail: those rules also beat some neural policies running on a graphics card. The field of application is narrow, and the direction matters more than the field. The large model is used once, to write the rules. After that, every request is handled by small, cheap code. Anyone signing a three-year data center contract today risks paying for capacity they will not need. The measure that counts is the cost per unit of useful work on the machines the client already owns. If the trend holds, the shift happens in 2028.

Good morning to everyone listening. The money moved on September 28, 2026, and it moved in exactly that direction. SiMa.ai closed a Series C round of 150 million dollars, at a valuation of 1.45 billion. Fidelity Management & Research Company and Amplify led the round together, and the San Jose company has now raised 500 million dollars in total. The comparison with the previous round measures the jump: in July 2025, after a Series B of 85 million, the price was 960 million, according to PitchBook data cited by TechCrunch. Fifteen months later the valuation grows by about 51 percent. The company makes chips and software to run AI models inside the machine itself: robots, drones, car cockpits. Krishna Rangasayee founded it in 2018, after serving as chief operating officer at Groq. The company says it will spend the money on two fronts: Palette Neat, the software environment for developers building on its platform, and new hardware expected in the first half of 2028, with one thousand dense TOPS of compute capacity.

Greetings to everyone listening. About 500 people are leaving Workday, and that figure equals 2.5 percent of an organization that counted more than 21,000 people at the end of January 2026. The number comes from the filing submitted on September 29, 2026, to market regulators, and the official reason speaks of aligning team structures with growth priorities, as Business Insider reports. The company says the cuts hit one team only, product and technology. Where the cuts land matters more than how many there are. Here they hit the people who build the product itself. This is the second such move of 2026. In February about 400 people left, and a week later the chief executive at the time, Carl Eschenbach, stepped down. In his place came back Aneel Bhusri, cofounder and executive chairman, now on his fourth turn leading the company. The stock is down about 12 percent since the start of the year.

Hi everyone. In September 2026 Target described how its website works during the back-to-school season. The wish list sits at the center of the page, next to a box that suggests what to buy next. The order of operations deserves attention: Target first measured how customers behave, and then built the model that follows that behavior. The company stated that customers who build a list generate about 45 percent more demand in the category, a figure reported by Retail Dive on September 3, 2026. The comparison behind that number is clear: customers who keep a list, against customers who do not, in the same category. That is what separates a real measurement from a marketing claim. The back-to-school season weighs heavily in the American retail calendar, with spending that could reach record levels this year, according to the same outlet. On numbers that large, an effect like this shows up in the company's accounts. Many companies choose the model first, then look for a place to put it. Target did the reverse, and the result is measurable.