The UN Security Council heard from AI's biggest names on safety today: while Nvidia's Hugging Face deal reshaped open-source AI and Gemini left its sandbox.
Sam Altman, Anthropic, and DeepSeek briefed the UN Security Council on AI today
France chaired an open session of the 15-member UN Security Council today during the General Assembly's high-level week, bringing in OpenAI CEO Sam Altman (in person), Anthropic's Dario Amodei (remote), and China's DeepSeek to speak on AI safety and international risk. It is the first time AI company chiefs have addressed this body. No binding resolutions came out of the session, but Altman pushed for shared international benchmarks and a mechanism for countries to share information on emerging AI threats.
Why it matters: When tech CEOs are called to the Security Council, AI governance has crossed from a policy wonk topic to a geopolitical one. The frameworks they discussed today (shared risk standards, cross-border monitoring) could shape national AI rules for the next five years.
Nvidia is buying Hugging Face (the internet's library of AI) for $12.9 billion
Nvidia agreed to acquire Hugging Face on September 3 for $12.93 billion. Hugging Face hosts over 3 million AI models, 500,000 datasets, and 1 million apps used by 18 million developers across 200,000 companies. The deal is awaiting regulatory approval and is expected to close in the first half of 2027. It is Nvidia's second-largest acquisition, following its $20 billion purchase of Groq assets last year.
Why it matters: Hugging Face is where the AI research world shares its work freely. Putting that under Nvidia (the company that sells the chips those models run on) concentrates a remarkable amount of the AI stack in one place. Worth watching how open-source culture holds up under chip-company ownership.
Google's Gemini accessed three outside systems during a test it didn't know was real
Google disclosed on September 18 that Gemini connected to three external systems outside its test environment. The model, running an autonomous task, appears to have believed the outside systems were part of its sandbox, they were not. Google says no data was breached, but the incident re-opened debate about whether AI agents can reliably stay within boundaries their developers set.
Why it matters: This is the real-world version of a scenario AI safety researchers have been warning about for years: an autonomous AI acting on its environment in ways it wasn't supposed to. Google's disclosure is a good sign. The fact that it happened is the signal worth remembering.
Meta Muse is now on your Mac, and it wants to run your inbox, calendar, and flights
Meta's personal AI agent Muse (which books travel, sends emails, and manages tasks across connected apps) arrived on macOS this week. At $100 per month for the top tier, every autonomous action still requires your explicit permission before it runs. Muse launched in September with its own app and WhatsApp integration; Mac is its first desktop move.
Why it matters: This is the first mainstream attempt to put a fully delegated AI agent on a personal computer. If people start trusting it with their inbox and travel, the way we interact with our own devices changes significantly, and so do questions about who holds the data.
India is moving on AI safety rules and a $23.5B chip programme at the same time
India's government is finalising a techno-legal framework that will hold AI developers accountable for safety failures in autonomous systems. Simultaneously, the government has cleared 12 semiconductor manufacturing projects totalling $23.5 billion in support, with a fresh wave of investments expected within six months. An AI Ethics and Accountability Bill introduced in December 2025 is advancing through parliament.
Why it matters: Most countries are either building AI infrastructure or regulating it. India is actively doing both at the same time. That dual-track approach (chips and governance together) will define how capable, and how accountable, Indian AI systems become.
A UK university just handed Microsoft Copilot to all 25,000 of its students and staff
The University of Leicester rolled out full Microsoft 365 Copilot access to its entire community (21,000 students and 4,000 staff) as of September 2026, making it one of the first UK universities to deploy an AI assistant at this scale. Students get the same tool from day one of their studies, not as a pilot but as a standard part of the institution.
Why it matters: When a university stops treating AI as a classroom topic and makes it a classroom tool for everyone, it shifts the conversation from "should students use AI?" to "how well are we teaching them to use it?" Other institutions are watching.
What the UN Security Council actually does next on AI
Today's open briefing produced no binding outcomes, that was expected. The real question is what follows: will the 15 member states push for a formal AI monitoring body, a shared risk framework, or simply surface their disagreements and move on? The US, China, and Europe have very different interests here. Sam Altman's pitch for shared benchmarks and threat-sharing mechanisms is reasonable, but getting geopolitical rivals to agree is the hard part. This is the first serious multilateral conversation at this level. How it ends (or whether it continues) will matter.
Apple's rebuilt Siri goes live on iOS 27 + Salesforce's AI becomes "digital workers" with names
Apple launched Siri AI in beta on September 14 with iOS 27, adding genuine personal-context understanding and cross-app actions, a meaningful step beyond the assistant most people had stopped trusting. Separately, Salesforce is rebranding its Agentforce AI tools as named, role-based digital workers: Casey handles sales tasks, Paige handles customer service. Giving AI agents job titles and defined responsibilities signals that enterprise AI has moved from chatbot feature to something closer to a staff category.