Transcript · 2026-10-22

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 October 22nd, 2026. Today, AI news is less about one new model than the systems around models: where companies can buy them, how developers might use them, and what happens when their capabilities find serious software flaws. First, OpenAI and Google Cloud say OpenAI's frontier models and Codex workflows are coming to Google Cloud's Model Garden and Marketplace in the coming weeks. That means Google Cloud customers will be able to use existing cloud spending commitments for the purchases. The companies say Google Cloud's access controls, data governance, and infrastructure are meant to support enterprise and regulated workloads. Reuters also reported the deal, which follows the end of OpenAI's exclusive Microsoft distribution arrangement in April. Microsoft remains OpenAI's primary cloud partner, and this is a planned general availability, not a service that is already broadly live. The practical takeaway is distribution. A company that already runs data and applications in Google Cloud may be able to evaluate OpenAI models in the environment it already uses, alongside other model choices. OpenAI also expanded ChatGPT ads to 26 additional countries, saying ads are now available in 89 countries and that more than 100,000 advertisers have used the product. The company says ads appear only on Free and Go plans, while Plus, Pro, and Enterprise remain ad-free. It also says ads are labeled, separated from answers, and do not influence the answers ChatGPT gives. An engineering post explains that ad selection runs in a separate service. A language model can interpret conversational context, retrieve candidates, apply eligibility and quality checks, and return no ad when there is not a useful fit. Those are OpenAI's product design commitments, not an independent privacy or relevance audit. Still, the technical detail matters because conversational advertising creates a different set of concerns from matching an ad to a single search query. Now to research. OpenAI, Perturb AI, and Scripps Research described using thousands of GPT Rosalind-powered agents to analyze existing data from millions of mouse brain cells. The team divided work across gene disruptions, cellular responses, literature search, and proposed explanations, then combined the results into what it calls a reasoning atlas. They say the analysis took 32 minutes. The researchers report that the agents found leads, including a possible connection between Dctn1 and cholesterol regulation. They also say blinded human experts preferred the distributed-agent reports over single-agent reports 75% of the time. But these are leads for further experiments, not a treatment result. The proposed biological mechanisms still need laboratory testing. The useful claim here is about expanding the set of questions scientists can investigate, not replacing validation. Security was another major theme. OpenAI's Patch the Planet initiative described coordinating around critical PHP vulnerabilities that it says GPT-5.6 Sol helped uncover, with Wiz independently reporting both vulnerabilities. According to OpenAI, PHP maintainers led the fixes, while researchers, Cloudflare, Wiz, and other partners helped test fixes, prepare guidance, and organize protections ahead of disclosure. For people running PHP, the immediate advice is straightforward: use the maintainers' official update guidance. OpenAI is also proposing a wider response model: validate and fix upstream, give affected organizations time to prepare, coordinate where government partners can help protect critical systems, prepare downstream updates, then publish guidance and support deployment. That framework is a proposal, not an adopted industry standard, but it focuses attention on a real operational question. When AI helps find a widely used software flaw, the hard part is not only proving it exists, it is getting a reliable fix to the right people before the risk spreads. Finally, Nvidia and Google Cloud announced A5X instances built around Nvidia Vera Rubin NVL72 systems, ConnectX-9 networking, Google Cloud's Virgo network, and liquid cooling. The companies say general availability is coming next month, followed by a planned expansion to Europe. Google Cloud also says it plans to scale Rubin-based infrastructure to as many as 960,000 GPUs across multiple sites. Those are forward-looking vendor plans, not yet demonstrated operating capacity. In the near term, the announcement is a new cloud configuration aimed at training, inference, and reinforcement learning workloads. Put together, today's stories describe AI becoming more embedded in the infrastructure and institutions around it. Enterprise clouds distribute model choice, research teams sift bigger datasets, security groups coordinate patches, and hardware providers plan the capacity underneath it all. That's Overnight AI for October 22nd, 2026. You can find the source links and the full briefing in the show notes.