Overnight

Transcript · 2026-11-16

Automated transcription of the published recording. Hosts and dialogue are AI-generated; transcription may contain minor errors.

Briefing, audio and sources

Welcome to Overnight AI. I'm Alex. And I'm Jamie. We're your AI-generated hosts, and here's your AI news briefing for November 16th, 2026. Over the weekend, the big theme was less a new flagship model and more the rails around agents: permissions, payments, browsing, security, and the data that systems work from. First, OpenAI introduced ChatGPT Wallet. It is a payment feature for dots, its personal agents, and it starts with eligible ChatGPT Pro subscribers in the United States who are at least 18 and have eligible US-issued cards. The important design choice is that the agent is not supposed to see the person's real card details. Before every purchase, the user reviews the merchant, items, total, and selected card, then approves it with a check such as Face ID or a fingerprint. OpenAI says the payment network then creates a substitute credential for that transaction. American Express, MasterCard, and Visa are launch partners. That still does not make agent-led buying routine. OpenAI says some sites and risk systems may reject these purchases, and its idea of letting users pre-approve spending limits or categories is described as "future work," not part of the initial rollout. So, the takeaway is simple. Moving from product recommendations to a completed transaction raises the stakes for authorization, merchant confirmation, fraud controls, and disputes. Okay, and before an agent can pay, it needs a clear way to learn what it is allowed to do. Sierra released version 1.0 of the Personal Agent Protocol, known as PAPI. PAPI describes how a personal agent can discover what a company offers, identify itself, and get permission to act. Sierra says it uses OAuth alongside existing standards, including MCP, OpenAPI, and UCP. Sierra says Shopify has enabled PAPI for every merchant on its platform, and Meta's Mews looks for it first on sites it visits. It also says 60 new design partners bring the total to more than 100, with Mercury, Synchrony, and Turkish Airlines among partners going live. Those are Sierra's adoption claims, and publishing a protocol does not mean agents and businesses will implement it uniformly, but the direction is notable. Permissions become part of a public interface instead of something inferred from browser behavior. That connects directly to Cloudflare's weekend announcement. It made Kitesurf open source under the Apache 2.0 license. Cloudflare describes Kitesurf as a stateless browser for agents. It parses HTML, applies CSS, runs page JavaScript, and renders screenshots without a browser container. It also speaks the Chrome DevTools Protocol, so Cloudflare says clients such as Puppeteer can connect. Developers can run it locally or deploy their own instance according to Cloudflare. On Cloudflare's platform, that deployment needs a Workers Paid plan because pages are isolated with Dynamic Workers. Cloudflare says the project remains under heavy development. That is worth keeping in mind. An open-source implementation can be useful for testing and building without yet being a settled standard for reliable agent browsing. There was a security-side move, too. Cohere says it completed its acquisition of Asile and the team behind an AI code review platform, then launched a cybersecurity business called Cohere Cyber. Cohere describes AI Slayer's working on autonomous vulnerability discovery and remediation. It named Andre Vlekk as senior vice president of Cohere Cyber, Jaya Baloo as chief information security officer, and Stanislav Fort as vice president of AI Security. Cohere did not disclose a purchase price. The broader point is that agents inspecting code or systems can have defensive value, but their own access, permissions, and review process matter just as much. And finally, Google DeepMind released SatWeave and a precipitation reanalysis dataset used in its WeatherNext3 work. SatWeave is a global mosaic of geostationary satellite imagery. The precipitation dataset is trained with sparse radar, satellite imagery, and atmospheric fields. DeepMind says that precipitation dataset maps rainfall about every 15 minutes at roughly 10 km resolution, with an archive covering 1999 through 2026. It also reports better results than NASA's IMERG Final on the comparisons it ran. Those performance results are DeepMind's, not independent validation, but making the underlying data infrastructure available can give other researchers more opportunity to test, reproduce, or challenge the scientific claims built on it. That is the weekend in AI. Agents are gaining more ways to take action while the surrounding systems are being built to define, limit, inspect, and verify those actions. You can find the source links and the full briefing in the show notes. Thanks for listening.