The AGORÀ Briefing — Wednesday, September 23, 2026

Good morning, Adam. The flaw is called CVE-2026-90898, and it carries a CVSS score of 9.8. JFrog Security Research disclosed it on September 22, 2026. Bifrost is a free and open program that routes a company's requests to more than twenty language model providers. A single anonymous web request is enough to make the server hosting the gateway run commands. A single anonymous request. The attack succeeds with no password, with no connection set up in advance between the two programs, and without anyone clicking or typing anything. Yuval Moravchick reconstructed the whole path. The attacker sends an anonymous request to the gateway's administration address and registers an external component. The gateway launches the named command immediately. The gateway runs the command first and checks who asked for it afterward. Every version before 2.1 that takes requests over the web is affected, whenever the protection on the administration interface stays switched off. That protection is switched off by default when the program is installed, as The Hacker News reports. And that program holds the access keys for every provider connected to it.

Hi Adam, the price list weighs a great deal. On September 22, 2026, Anthropic released Opus 5.5. The number that moves decisions is the price. Output tokens cost 20 dollars per million against 25 for the previous model, as TechCrunch reports. Opus 5 had arrived on July 24, 2026. Two months later, the priciest model in the Claude line costs 20 percent less and, according to the company, beats the largest Fable model in many direct comparison tests. On the commercial side, what matters is the cost of running the model. Opus 5.5 answers faster and uses less computing power. So the price cut stays. It is the new list price, good from here on. Companies that signed a fixed-price contract in 2025 for the use of these models find themselves, within three months, paying far above the current price. The most powerful models on the market keep getting cheaper, month after month.

Hello everyone. A comparison like that holds up less than it promises. PsyAgentBench, filed on arXiv on July 23, 2026, sets two numbers side by side: 0 percent and 83.3 percent. The experiment is always the same Asch conformity study, and the model is always the same one, gpt-oss-120B. The single thing that changes is the instruction sentence given to the model. That alone. When the test reaches the model as an ordinary task, the model keeps its own answer every time: zero cases of falling in with the group. When the instructions say openly that this is that famous test, the model falls in with the group in more than four cases out of five. The author, Joy Bose, tries four combinations. Each test reaches the model either under its own name or as an ordinary task. And each test is used either in the classic textbook version or in a rewritten version, so the model has not already seen that exact wording in its training texts. The model, the task and the training were the same, and the result moved by 83 points. A jump that big says the result depends on how you ask the question, more than on the model.

Good to be back with you. The brain of the humanoid robot is by now a standard product, the same for everybody. The facts show it. On September 21, 2026, Hanxiao Chen published on arXiv an LLM-based conversational assistant for MyBuddy, a 13-axis humanoid driven by a Raspberry Pi. The hardware and software together recognize speech in real time, reason about complex questions, search the web on sources like Wikipedia, hold up a dialogue, and answer in a natural voice. The work was accepted as a poster at the WiML workshop at NeurIPS 2026. A desktop robot today sustains a long conversation and remembers what was said a few exchanges back. The hard part now lies elsewhere: the motors that move the robot, the hand that grips objects, and the cost of every working hour of the machine. Anyone writing a business plan that puts intelligence among the biggest cost lines has the math wrong. The expensive piece is the robot's body.

Greetings, Adam. The decisions are made in supply contracts, before any treaty. In June 2026, Alibaba Cloud opened two facilities in Johor, bringing its data centers in Malaysia to five. That is the group's widest presence in Southeast Asia, as The Diplomat documents. The opening came four months after ASEAN signed off on an agreement on data crossing borders, an agreement pushed by Malaysia. It sets shared principles on how data gets protected, on what the authorities may see, and on where data may be stored. Those shared rules will be rewritten several times over the years. Those machines will sit there for twenty years. The shared ASEAN rules set general principles. The contract written by the service provider sets precise obligations. In 1853, the first Indian train covered the thirty-four kilometers between Bombay and Thane. The money was British, and the British also chose the width of the track: 1,676 millimeters, the width India still uses today.

Morning, Adam. The California Court of Appeal decided the case of Del Biaggio against Bansen. The first public analysis of the ruling carries the date of September 22, 2026. The case grows out of a personal services contract. Daniel Del Biaggio worked four years at Bancrest Dairy and sued Pete and Mary Bansen for breach. The jury awarded 52,850 dollars. The defense, which won the case, then asked for 115,533 dollars covering attorney fees and paralegal hours. The trial judge trimmed the attorney hours and excluded the paralegal hours entirely. On appeal, the fee cut is upheld for three reasons: the hours were billed in a block with no detail, some work had been done twice, and other work concerned claims the judge had rejected. The paralegal hours, on the other hand, must be paid. The second part of the ruling concerns conduct in the case. The attorney had filed citations to rulings invented by artificial intelligence, and a sanction followed.

A good day to everyone listening. A Talogy study published on September 22, 2026, gathers answers from more than 200 executives who handle personnel, hiring, and training. 78 percent say they are worried. Tomorrow's leaders are learning less about leading, because the tasks of entry-level roles are passing to artificial intelligence systems. The people answering decide who gets hired, who moves up inside the company, and who will take the leaders' place. They see the problem from inside, because they are the very ones building those paths. Fewer people are starting out on the path toward management jobs, so fewer of them are learning how to lead. HR Dive records the same worry across every department surveyed. For decades, managers trained themselves by doing the everyday work: small, clearly defined tasks, where a mistake only affected a minor case. You learned the craft by getting things wrong. That daily apprenticeship is fading away. All the same, we still need someone able to spot that a system's protection stayed switched off, exactly as it came out of the factory.