Transcript · 2026-10-15

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 15th, 2026. Today, the thread is where agents are being put to work, and what institutions say they need around them. First, OpenAI published an analysis of agent use in European businesses. It says weekly active users of Codex and ChatGPT Work at small and medium-sized EU businesses reached 8.1 times their May level. That sounds like rapid adoption. What exactly is being measured? OpenAI is looking at its ChatGPT Enterprise and business accounts, not every European business. In that sample, agentic tools generated more than half of EU enterprise output tokens. So, it is evidence about use of OpenAI products, not proof that companies are already more productive. The top 10th of EU firms by output tokens per active user produced more than four times as many tokens as typical EU firms in September. Leading firms outside the EU were still more than twice as intensive as those EU leaders. The useful takeaway is that the move from asking a model questions to assigning multi-step work is showing up in usage data. But the benefits outside that data set still need independent measurement. Next, Singapore was a focus for two separate announcements. Anthropic opened its fifth Asia-Pacific office there, and signed a memorandum of understanding with Singapore's Ministry of Digital Development and Information. What does the memorandum actually commit the two sides to? Anthropic says it plans cooperation on model evaluation, safety and security, adoption research, skills, and responsible use. It also says it will exchange information with Singapore institutions on testing standards and incident response. That is a framework for planned collaboration, though not a result from a new safety program. Anthropic also points to local government and commercial use cases, but those operational benefits are its claims. The near-term signal is the local office and the agreement to exchange technical and policy information. And Google DeepMind had a related Singapore announcement, but with a much narrower application. DeepMind and Singapore's GovTech announced a scam detection model for typical smartphones. They say a lightweight Gemma model was fine-tuned on anonymized local scam material, and can warn about manipulative conversational patterns. The on-device part matters because encrypted messaging services cannot simply inspect private chats in the cloud. That is the stated design goal. Google says the model processes conversation data on the device, and that training excluded citizens' private or personally identifiable data. The project was presented at Singapore International Cyber Week. Do we know whether it works well enough for a broad deployment? Not from this announcement. It gives no false positive rate, coverage figures, or deployment timetable. It describes a two-stage pipeline and local testing, but those questions matter before a warning system can be judged. Now to the UK, where Nvidia announced a large compute commitment for safety research. Nvidia says it will provide a five-year compute donation, valued at $250 million, for UK agentic safety and security research: $50 million in access each year through UK-based Nvidia cloud partners. And some of that access is directed to named research groups? Most capacity will support the wider UK research community, with a portion for the British Open-Ended Learning and Discovery Lab, the Science of Fundamental AI Research Lab, and Imperial College London. It is important not to turn that into an evaluation result. This is a pledge of computing access, not a cash grant or evidence that a particular safeguard works. Its value will depend on allocation, access, and published findings. But it is a concrete resource for testing systems that can plan and act across longer tasks. Finally, a reported benchmark puts a different kind of pressure on safeguards after a model has been released. Wired reported on a counter-terrorism AI benchmark by UK non-profit Tech Against Terrorism. It said the benchmark tested 134 models with more than 200 dangerous prompts, and only two passed its full threshold. It also found that framing a request as safety research made models more likely to answer, according to the reporting. Wired further reported that stripped safeguard versions of new open-weight models now appear online in less than three days on average, versus more than 24 days in 2024. Those are reported findings, not independently reproduced results. Still, they describe a practical limit on a developer's control. Once downloadable weights are modified and redistributed, the original refusal behavior is no longer the only safety measure that matters. That is the common thread today. Adoption measurements, public sector partnerships, and compute pledges all concern what happens when agents leave a lab. The questions are practical: what is measured, what is tested locally, who can intervene, and what remains true after release? You can read the full briefing and its source links at overnight.cognitos.com. Thanks for listening.