The AGORÀ Briefing — Thursday, September 3, 2026

Good morning Adam, good morning everyone. The bug is called CVE-2026-82329 and carries a CVSS score of 9.8. JFrog, the company behind Artifactory, made it public on September 1, 2026 and shipped the patch the same day. The mechanism: an anonymous request, sent to an exposed endpoint with zero identity validation, turns into a full-access administrative token. The Register documented the first active exploitation within four days. The team at the security firm watchTowr saw attackers generating admin tokens on its honeypots, from IP addresses spread across different geographies. Yordan Ganchev, its principal threat intelligence specialist, describes systematic enumeration of users, groups and federated credentials, and warns that wide-scale scanning remains a likely scenario in the coming days. The Register's headline reads "Another Artifactory CVE": that word, another, tells a story of recurrence. A breach like this reaches AI agents too, and anyone with an exposed instance should update today.

Hello Adam, hello to everyone listening. Agents on power grids finally have a public yardstick. On August 31, 2026 six researchers posted a benchmark on arXiv called RestoreBench, reference arXiv:2609.00384v1. It evaluates agents based on Large Language Models on one precise task: restoring power flow convergence in electrical grids. It covers two grids and 46 cases per grid, for a total of 92 failed convergence scenarios. It compares three architectures, chatbot, single agent and multi-agent, in the same simulation environment and with the same metrics. That covers capability. The real transformation concerns accountability. A reproducible benchmark makes an agent's performance falsifiable, and therefore makes regulable what used to remain engineering opinion. The question shifts from can we deploy to who decides. Grid operators now hold a shared yardstick, and from that yardstick the serious confrontation with the EU AI Act begins.

Good to see you all, and good morning Adam. Before the benchmark, the precedent, which sits in the year 2000. Christina and David Romer, in the American Economic Review, showed that the Federal Reserve's internal forecasts, the so-called Greenbook, systematically outperformed those of private analysts. Resources, access, proprietary data: the central bank knew first and the market chased. That gap sustained the architecture of monetary decision-making for forty years, and today it erodes. A language model with web search replicates what once required a department of twenty economists. The evidence dates to August 31, 2026: six researchers published "Can LLMs Take the Pulse of the Economy?" on arXiv and built LiveMacroEval, a contamination-resistant benchmark. The scarcity of macro information, controlled by policy makers, becomes a good available to anyone who pays for a query. This is a regime change, and it changes the math for anyone managing capital.

Hi Adam, and greetings to everyone tuned in. In the raw data, exactly, and that is precisely what the market misprices. In materials science, value migrates from the AI model to the dataset that trains it: that is the documented trajectory of recent years. Consensus funds algorithms. The real leverage lives in extractable scientific literature: whoever controls that corpus dictates the timing of a generation of products. The proof comes from Seoul. On August 31, 2026 a team at Seoul National University, led by professor Ho Won Jang, made its result public. The researchers mined 1,202 records from 448 papers and, with physics-informed machine learning, screened about 150 million virtual compositions of lead-free dielectrics. The filter narrowed the field to 37 candidates. Two were synthesized and tested, both with high dielectric constants and strong stability at elevated temperatures. These are the materials of the MLCCs inside smartphones and electric vehicle electronics. High confidence on the technology, medium on the timing: a 24 month horizon.

Welcome back to the table, Adam, and good morning to everyone listening. The evidence comes from an interpretability study published on September 1, 2026, which compares three modes of human-AI interaction in learning: an unrestricted conversational chatbot, a Socratic mode that guides through hints, and an adaptive tutoring system that adjusts difficulty in real time based on the brain signal picked up by a Muse EEG device. The unrestricted chatbot group achieved higher learning gains than the other two conditions, with p below 0.03 and d above 0.80. The adaptive condition instead generated significantly higher EEG engagement, with p equal to 0.018. More brain effort on one side, more score on the other. And the question that matters for anyone allocating budgets remains: does that gain, measured right after use, reflect acquired competence, or is it an artifact of the measurement time window? The study adds evidence, and the question stays on the table.

Good day to everyone, Adam included. The bet has a precise number: on September 2, 2026 Workday announced that over 100 U.S. state and local agencies have chosen its unified platform in the last two years. The list includes the State of Delaware, the Commonwealth of Massachusetts, the New Jersey Transit Authority and the New York State Unified Court System. The cadence tells the dynamic: a public agency signs almost every week, and a rhythm like that signals structural pressure, far more than a technology fad. Michael Hofherr, the company's senior vice president of industry product, sums up the stakes: public servants deserve reliable technology to hire, to pay people and to read the state of a budget. The problem is operational before it is technical: many administrations deliver essential services on fragmented systems, with limited time and resources. Technology pays off when people know how to use it, and the competence of those people is the real investment.

Good morning Adam, and good morning to everyone following us. Behind it is a move across three markets, and the story begins on September 1, 2026. GoPro announced its merger with Starman Optical for 285 million dollars, with a declared goal: maximizing the value of its intellectual property across consumer, commercial and defense. TechCrunch reports that the deal pays shareholders 1.14 dollars per share and leaves them with about 10 percent of the new combined entity. It retires 92 million dollars of outstanding debt. Closing is expected by the end of the year, and GoPro remains a publicly traded company with its shareholders on board. For a decade the market valued GoPro as a maker of action cameras. Starman Holdings values it as a holder of precision optics expertise. Charles Tebele, its CEO, states it explicitly: advanced optics and imaging are essential for AI and for national security. The action camera stays in play, while the real value sits in the optics.