The AGORÀ Briefing — Saturday, October 3, 2026

Good morning, everyone. CVE-2026-90970 carries a CVSS score of 9.9 out of 10 and it hits GitLab's AI Gateway, the service that connects a GitLab instance to AI models. The company disclosed it on October 2, 2026, with an advisory that classifies it as critical. According to the advisory, picked up by The Hacker News, a user who is already authenticated and has access to the Duo Agent Platform can run commands on the gateway. There are three fixed versions: 19.2.4, 19.3.2 and 19.4.1. GitLab.com, GitLab Dedicated and self-managed instances that use a gateway hosted by GitLab are already in the clear. Anyone hosting the gateway inside their own infrastructure stays exposed. GitLab sent those customers the warning privately, ahead of publishing it. The recommendation is to update right away.

Good morning, Adam. Three researchers compared two ways of delegating work to agents across 77 scenarios and five frontier models, in a study posted to arXiv on September 30, 2026. A team of agents, each one serving a different user, produces worse group outcomes than a single agent serving everybody. The gap shows up in all four environments they measured. The design keeps the shared resource fixed and changes who governs it: an API key with a shared compute budget, a clinic with a shared calendar, a personal assistant with a group order, a merge queue with a release deadline. In the personal assistant environment the single coordinator satisfies a targeted request about twice as often as the team does. The work is signed by Sahan Paliskara, Nattaput Namchittai and Andrew Lampinen.

Hi Adam, and here there are two dates to mark. On September 30, 2026 Reddit announced that RSS feeds stop on November 13, 2026 and that the public API closes by March 2027. The stated cause is automated scraping by bots that harvest text to train models. In the announcement the company writes that RSS became, quote, a common surface for large-scale scraping and automated abuse, as TechCrunch reported on September 30, 2026. RSS lets an app read a site's updates in a standard format. It's the tool that made blogs popular, and Google ran the most widely used reader for it until 2013. What matters here is who controls the channel that carries the text. To moderators, Reddit points at the Discord Relay app as the replacement. Text written by people becomes a commodity you buy, with a supplier, a price and a contract.

Good morning. On September 29, 2026 the President of the United States signed executive order number 14434, titled "Inaugurating the Era of Super Intelligence." The act appears among the presidential documents published by the White House. The text orders one thing only: a change of words. Federal executive branch bodies adopt "Super Intelligence" and "SI" as their working terms, and the words "Artificial Intelligence" and "AI" lose standing in the administration's official documents. The order number and the signing date become the official citation once the act enters the Federal Register, the federal archive of presidential documents. From that moment on, every rule that invokes the order cites that number and that date. The reach stays administrative and internal to the executive: the word changes, and every other rule stays as it was.

Good morning to you and to everyone listening. In 2026 Omni Calculator surveyed 705 American workers who use AI to write, and it found an inverted pattern of checking: the higher the rank, the more the text gets reviewed. 54 percent of managers and 48 percent of executives run at least 70 percent of what they write through AI before sending it, against 30 percent of senior individual contributors, according to the report picked up on September 30, 2026. The authors call it the managerial paradox: automated verification climbs toward the top, where the approvers sit. The sample stays narrow: people employed in the United States who already use AI for writing. 81 percent of executives and 78 percent of managers say AI changed whether they send a message at all, and how they word it. How ready the leadership is shows up in one tiny gesture: who checks the text before sending it.

Good morning, and I'll start with a deadline. The premium paid today for HBM4 memory disappears by 2028, while analysts treat it as locked in through 2030. That leaves two years in which those high margins still hold. In the quarter closed on September 3, 2026 Micron turned 69.5 percent of revenue into net income: 54.23 billion dollars in sales, 43.75 billion in operating income, 37.7 billion in net income, as The Next Platform reported on October 2, 2026. Analysts read that margin as proof of a technical edge that is hard to copy, because stacking memory chips and bonding them to the compute chip remains a craft for a handful of plants worldwide. The number that sets that deadline is what Micron spends on new plants: 10.77 billion in capital spending against 73.45 billion in cash. With that cash Micron has the money for almost four new fabs, and it commits a fraction of it. What limits production is the physics of making these chips, not the money available.

Good morning. On October 2, 2026 the ECB blog published the first broad measure of how European firms intend to pay for artificial intelligence. The source is the Survey on the Access to Finance of Enterprises, around 5,000 euro area firms, with questions about the twelve months ahead. In West Germany in the fifties, industrial reconstruction was paid for in large part out of retained earnings: margins were high, dividends were low, and the tax code rewarded whoever reinvested. In Japan after 1990 the same pattern came back inverted, with cash going to pay down debt and the central bank holding the cost of money near zero. In both cases the central bank learned the same thing: changing interest rates only moves company spending when that spending is funded by borrowing. Two precedents suggest a rule, and a third case would confirm it. That third case is Europe today.

Good morning, and the story starts in a lab. Takeda Pharmaceutical brought AI into its research and development, and the scientists changed how they work: instead of running broad experiments, they let the model predict from data, then test only a few hypotheses. The case came out on October first, 2026 in MIT Sloan Management Review, written by George Westerman and David Kiron. It's part of a program the magazine runs with EY, which funds it. Keep that in mind when you read the numbers. Takeda is headquartered in Tokyo, with two large hubs in Zurich and in Cambridge, Massachusetts, inside a group that grew by acquisition. What made the difference here is how they redesigned the work, more than which technology they chose. The scientist's craft stays at the center: the model cuts the number of hypotheses, and the person decides which one earns a test at the bench.