Transcript · 2026-10-28

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 28th, 2026. Today, OpenAI and Oracle are planning tighter enterprise integrations, Anthropic has published unusual detail on cyber research and expansion into government and drug development, and Meta is bringing its Muse agent to Singapore. First, OpenAI and Oracle say they are expanding their partnership to bring Oracle applications and data into ChatGPT. Their planned Oracle AI Workspace is meant to let employees find information and complete approved workflows across finance, procurement, and planning systems while retaining their organization's permissions and approvals. That is a meaningful constraint. The pitch is not simply that a chatbot sees more company data; the intended setup connects work across business software, but says access should still be governed by each customer's sign-in, permissions, and approval process. The partners also announce an AI Security Mesh for finding and fixing vulnerabilities in authorized environments, an SMB bundle that combines ChatGPT Business and Codex with Oracle products and cloud credits, and an app composer for building internal apps against approved Oracle data and workflows. Oracle customers in the United States will be able to buy the OpenAI API, Codex, and ChatGPT Work through Oracle Marketplace. OpenAI says Oracle reported 130,000 active ChatGPT users and more than 95,000 active Codex users. Those are the partners' figures, and the useful test will be how reliably the integrations work once customers configure them. Security is also at the center of Anthropic's new research disclosure. The company says an unreleased internal research model helped find five attack chains in widely used browsers and operating systems, including a one-click Chrome remote code execution chain, a Safari chain, an Android lock screen bypass, a Windows denial-of-service path, and a Linux zero-click chain. Anthropic says the research involved sophisticated, multi-step exploits in both open-source and proprietary software, sometimes with little or no user interaction. It also says it reported the findings through coordinated disclosure, which gives vendors 90 days to patch and another 45 days before further discussion beyond public advisories. We should be precise about what that establishes. This is Anthropic's account of internal research, not an independent measurement of model capability, and the company is withholding operational detail for safety reasons. But it does make vulnerability research a concrete capability and risk test, rather than an abstract concern about frontier systems. A related question is how those systems enter public institutions. Anthropic announced a General Services Administration agreement that it says gives U.S. federal, state, local, and tribal government buyers a 40% discount on Claude Enterprise, Claude for Government, and the Claude API. Anthropic says the discount applies to direct purchases now and through cloud providers in coming months, with no minimum spend. It also says around 500 U.S. government organizations use Claude, and estimates the agreement could save buyers $1 billion. Those adoption and savings figures are vendor claims, not independently audited outcomes. Still, the agreement moves model purchasing and implementation deeper into normal public sector procurement. Anthropic is also launching a program to help agencies put the tools into use. That support may be as important as a price reduction, because permissions, review, and operational practices determine what an agency can responsibly deploy. Anthropic also gave an update on drug development work. It says it has partnered with BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, and The Column Group, while starting some drug discovery efforts of its own. The company describes Claude helping with tasks from target research and molecule design to manufacturing planning, regulatory filings, and clinical trial work. One reported Generate Biomedicines case used Claude Code to propose a shorter genetic design for a CAR-T cell treatment. Anthropic says lab tests showed stronger receptor production and more effective tumor cell killing. But the qualification is important. Anthropic says most of the work remains preclinical, meaning the drugs are generally not being tested in people yet. These are early research programs and company-reported results, not proof of a new therapy or evidence that AI has solved clinical development. Finally, Meta says Muse, its personal AI agent, is now available in Singapore on iOS, Android, and the web. Users can connect Gmail, Google Calendar, and Instagram. Local partners include Shopee, Grab, Singapore Airlines, Agoda, ShopBack, Klook, Trip.com, and Lazada, though some connector experiences will arrive progressively. Meta says Muse uses a dedicated cloud environment called Muse SecureVM to store connected service data and credentials, while keeping passwords and payment methods out of the agent's view. It says the user chooses what the agent can access, and is asked for sensitive actions, such as sending an email or making a purchase. Those are Meta's design and security claims. The harder, longer-running question is whether people can understand and manage those choices as more services, purchases, schedules, and messages become connected to an agent. Permission prompts are useful only if people can make informed decisions with them. The throughline today is integration with boundaries attached. Business systems, vulnerability research, government procurement, life sciences work, and consumer agents all depend on access rules and review, not just model capability. For the original announcements and their qualifications, find the source links in the show notes. Thanks for listening to Overnight AI.