The AGORÀ Briefing — Wednesday, October 14, 2026

Good morning, everyone. On October 1, 2026, the FDA published its guidance priorities for fiscal year 2027, as Healthcare Dive reports. Regulatory Focus from RAPS also covered the agenda. The document comes from CDRH, the FDA center that regulates medical devices. The top tier includes the final guidance on how companies apply to sell AI-enabled medical devices and how they manage those devices over their life cycle. The priority lists include seven final guidances and one draft. The draft covers the evidence companies must provide for generative AI chatbots used for mental health disorders. That topic was missing from last year's list. It comes after the FDA asked for public comments in August on how to regulate generative AI.

Good to be with you, Adam. A Stanford team read a quarter of a million anonymous conversations between people and Claude, the chatbot from Anthropic. The study has the backing of the Stanford Institute for Human-Centered AI. The journal Nature Human Behaviour has accepted it for publication. The conversations come from real users working on real tasks. The researchers analyzed them with Insights, the analysis tool from Anthropic. They measured three things: how serious each task is, how much control the person keeps, and how much friction there is between the person and the AI. Diyi Yang, a computer science professor at Stanford and a senior fellow at HAI, leads the team. The first authors are Yijia Shao and Dora Zhao. According to the Stanford HAI announcement, 56 percent of the conversations involve tasks with serious consequences that are hard to undo, or even more serious ones.

Good morning, Adam. LiteLLM is an open source project born at Y Combinator. Its gateway sits between applications and model providers. It holds the API keys, decides where each request goes, and keeps the logs and the usage limits. On October 13, Check Point Research published an analysis of version 1.77.7. The endpoint that updates the configuration lacked a permissions check. So a user with any ordinary account, with no administrator role at all, could rewrite the gateway's live configuration. With that change, according to the researchers, that ordinary user could read files on the server, take over the administrator account, and run code remotely. The flaws are tracked as CVE-2026-35029 and CVE-2026-59823. The report mentions no cases of anyone actually exploiting the flaws. Version 1.83.9 fixes both flaws.

Hello to everyone listening. On October 14, LG CNS announced a memorandum of understanding with KDDI, the Japanese telecom group. The goal is to sell artificial intelligence and digital services to companies across Asia-Pacific through the KDDI network. According to the Korea Herald, KDDI has 19 offices in 15 countries in the region. LG CNS brings two platforms. AgenticWorks offers custom AI agents for local companies. PhysicalWorks automates factories, logistics and retail, and it manages robots. A joint working group is looking for smart city projects in Vietnam, Thailand, the Philippines, Indonesia and China. The Chosun Ilbo reports that Park Sang-kyun of LG CNS promises "concrete results." The memorandum announced on October 14 contains no figures.

Hello, everyone. International Workplace Group, IWG, published the Z-Suite Report, which HCAMag picked up on October 12. According to the report, 92 percent of senior leaders would appoint a Gen Z person to a top role. For 82 percent of those same leaders, artificial intelligence speeds up the rise of younger people. The report surveyed two groups. On one side, senior leaders. On the other side, young leaders and people between 21 and 29. The HCAMag article leaves out the sample size, the collection method and the countries involved. So those percentages describe an intention. Saying "I would appoint" costs little. Putting a 25-year-old in charge of a whole department takes evidence of that person's judgment.

A good morning to all of you. The limit lies in household demand. On October 12, the Bank for International Settlements published Working Paper 1381 by Phurichai Rungcharoenkitkul. According to the paper, the boom in artificial intelligence investment hits its limit in how much households are able to buy. Chips and energy come after that. In the scenario the paper considers highly plausible, growth speeds up for a short time and then falls below 2 percent, the growth rate before AI. The AI industry sees the limits somewhere else. CoreWeave, which sells computing power for AI, devotes a blog post to the bottlenecks slowing AI growth: chips, energy and available GPUs. In 1820, Thomas Malthus asked David Ricardo who would buy the goods produced by capital that grows faster than consumption. Two centuries later, that question returns in the paper from the Bank for International Settlements.

Have a good day, everyone listening. Kogan is an Australian online retailer, and its customer support had always worked through web forms and email. CTO Goran Stefkovski saw AI as a way to make things easier for customers and to answer them faster. The first step was a deterministic bot that followed handwritten rules. Stefkovski describes it as "the technology we had available." That bot resolved about 6 percent of requests. Kogan kept the project but replaced the technology behind it. It rebuilt the bot as an AI agent on Salesforce Agentforce, with forward-deployed engineers, the technicians Salesforce sends inside its client companies. It took six weeks to go from the first trial to full production. Today the agent handles 10 types of request, as Stefkovski told the Dreamforce conference, according to iTnews on October 14.