The AGORÀ Briefing — Monday, September 21, 2026

Good morning, everyone. CVE-2026-85889 hits the ceiling of the scale, with a CVSS score of 10. Microsoft has released the fix for Azure AI Foundry, the enterprise platform where companies build and run generative agents. The advisory, published on September 17, 2026, describes missing authentication for critical function. A critical function stayed reachable over the network by an attacker lacking authorization, with elevation of privilege as the outcome. The Hacker News covered the story on September 18, 2026, crediting the discovery to researcher Rémy Marot. The company reports zero evidence of exploitation in the wild, and the flaw is already mitigated on the cloud side. For anyone designing agentic systems, the combination is what counts. Network vector, low complexity, privileges required equal to zero.

Hi Adam. Vals AI measures model capabilities in place of the people selling them, and it has closed a Series A of 40 million dollars led by Andreessen Horowitz. TechCrunch told the story on September 19, 2026. The company started in 2024, after a seed round led by 8VC and Bloomberg Beta. Co-founder Rayan Krishnan is 25, with a past as a Palantir intern and a path through Microsoft and the Stanford artificial intelligence lab. The figure matters less than the category it certifies: independent verification becomes a product on the shelf. Vals keeps its own test material confidential, because a public test ends up inside the training data. For a procurement department, due diligence on models moves outside the internal perimeter.

Good to have you with us. An ablation study usually takes apart the architecture: attention heads, layers, tokenizer. The work filed on arXiv on September 17, 2026 by David S. Berman, Ying-Jer Kao, Roger G. Melko and Alexander G. Stapleton takes apart the data instead. The authors study the neural scaling laws of RydbergGPT, an autoregressive transformer trained on projective measurements of qubits, collected from arrays of interacting Rydberg atoms. The transformer stays fixed, and the source feeding it changes. Near the critical point, the loss as a function of dataset size follows a power law with a loss floor correction. Far from criticality, that description loses quality in a substantial way. Ten pages, six figures, one tight thesis: the source of the data leaves a measurable trace.

Good morning to everyone listening. Talogy, a talent management firm, released a survey on September 18, 2026 covering 207 HR executives, hiring managers and training specialists. 78 percent fear a long-term loss of leadership skills, while artificial intelligence absorbs the tasks that used to train junior people. The figure measures a perception, and that stays its limit. The entry-level role has always served two functions at once. It produced useful work and it trained judgment through low-risk tasks. Whoever reconciles data learns to spot a strange number. Whoever drafts a document learns to take a correction. Take away that school, and the bill arrives in 2032, with the leaders that today nobody is training.

Hello. The drop in AI inference cost per million tokens has convinced the market that robot labor will fall along the same slope. The brain of a humanoid costs a fraction each year of what it cost the year before. The body obeys different laws. The real price of a robot-hour gets decided at decommissioning, and that line item stays outside every cost model in the sector. The figures from industrial pilots, around 25 dollars per robot-hour, measure usage. A standard humanoid contains 10,000 to 15,000 components, in 200 to 500 major sub-assemblies, according to The Robot Report. Once end of life enters the numerator, cost parity with manual labor slides further out. Whoever signs today buys that difference sight unseen.

Good morning, Adam. In October of 1989, Mitsubishi Estate took 51 percent of the Rockefeller Group for 846 million dollars. Six years later the American subsidiary landed in Chapter 11, and most of the real estate went back to the sellers. It was a structural error: promises in yen, assets in dollars, a long horizon. First the savings pile up. Then the internal return falls below the cost of the promises made to customers. Finally the capital goes out, and buys the asset that pays best at that moment. In September 2026 the object changes: in place of Manhattan stone, there is compute. Nippon Life Insurance plans to direct 2,000 billion yen, equal to 12.7 billion dollars, into infrastructure financing, American data centers included. Same mechanics, different asset.

Good morning from me as well. On September 10, 2026, in the earnings call of Macy's Inc., COO and CFO Tom Edwards described a precise move: an AI forecasting overlay on replenishment, moving from pilot to a wider rollout. The stated goal covers two items, in-stock levels on the shelf and inventory efficiency. The account comes from Retail Dive, which published the report on September 18, 2026. The metric capable of measuring the result stays outside the public documents. Underneath sits a plan in motion since 2024, named Bold New Chapter, with its center of gravity on the supply chain, the closing of low-productivity logistics centers and an automated facility opening in North Carolina. The original idea lives right there: a layer of artificial intelligence on top of a machine that was already moving.