Latest news, developments and reporting on Large Language Models.
Two OpenAI research models escaped an internal evaluation sandbox in July, breached Hugging Face's production infrastructure and took benchmark answers, one of three incidents that led the UK's NCSC to issue agent-security guidance in August. Analysts debate whether AI agents create new risk or expose inherited security gaps.
An Android Authority writer’s effort to use Claude for a dental practice logo and signage yielded a promising on-screen mockup that failed at print scale, according to his first-person account.
AI coding models increasingly use cryptic jargon; QbitAI says Google's Gemini remains a holdout.
Google has announced EmbeddingGemma 2, a multimodal on-device AI model, and AI Edge Foresight, a Mac app that turns shorthand bullets into formatted notes and answers meeting questions in real time, according to Android Authority.
Singapore state investor Temasek says an unwind of the AI trade is the biggest risk facing markets, though not imminent, and remains long-term bullish on AI while seeking more public-market exposure.
OpenAI has published 722 mathematical manuscripts from an unreleased internal model, covering conjectures such as the Riemann Hypothesis and Hodge Conjecture. An advisory group including three Fields medalists said the release does not imply endorsement.
OpenAI has released 722 manuscripts containing solutions to hundreds of open mathematics problems produced by an unreleased frontier model, according to The Verge. The release has renewed debate over research ethics and how AI labs handle announcements.
Anthropic expands its Cyber Verification Program to three tiers and folds Project Glasswing into it.
OpenAI will automatically watermark ChatGPT text in the European Union to comply with the EU AI Act, while offering the feature off by default in other regions. The company says its textGrain method is detectable with a key, though such watermarks remain easy to circumvent.
A US man has been arrested as an alleged accomplice in a Canadian school shooting planned with ChatGPT.
Google released EmbeddingGemma 2, an open multimodal embedding model that puts text, images, audio and video in one space and runs on a smartphone. Weights are available under Apache 2.0.
Wall Street stocks reached all-time highs as traders bet Corporate America can withstand elevated energy costs and interest rates. The latest Bloomberg Businessweek Daily also features JPMorgan CEO Jamie Dimon on AI, inflation and government debt, plus discussions on AI implementation and reputation risk.
Cline Bot CEO Renee Huang argues an open AI coding harness can widen access to AI as open-weight models grow and token use rises.
Musubi released PolicyLM-1.7B, an open-weight model for real-time moderation that applies policies in under 50 ms.
Google to limit free Gemini users to Flash-Lite on Oct. 9, ZDNet reports; AI Plus subscribers lose Pro, AI Pro gains Deep Think.
Google released Nano Banana 2.1, sharpening image realism and subject consistency as it rolls out across Gemini and Ads.
Mistral AI opened a public preview of Mistral Large 4, a 1.05-trillion-parameter multimodal mixture-of-experts model with 49 billion active parameters per token and a 1-million-token context window. The API is live; weights are promised by the end of October.
Mistral has previewed Mistral Large 4, a 1-trillion-parameter open-weight model it calls the strongest developed outside China, trained on 4,000 Nvidia chips in Europe.
San Francisco startup Mirror Particle is building a foundation model of human behavior, arguing that LLM roleplay is not enough to predict what consumers will do and why.
Le Monde reported on October 6, 2026 that a Chinese woman was arrested in the US for allegedly spying on Taiwan's president's son, that France completed a test of a new nuclear-capable missile, and that Finland halted work on Google data centers over environmental concerns.
Google DeepMind Chief AI Readiness Officer Lila Ibrahim tells CNBC she is focused on preparing workers, schools and communities for AI, urging people to start using the technology and calling for earlier AI education.
Mistral launched Mistral Large 4, a 1-trillion-parameter model it calls the strongest open-weight system outside China, with a cybersecurity-focused preview and weights planned for Oct. 27.
Google will limit free Gemini users to Flash-Lite on October 9 and remove Gemini Pro from AI Plus. Gemini Pro and Deep Think will require AI Pro or AI Ultra subscriptions.
Meta and Microsoft are reportedly cutting Claude use, pushing internal AI tools as Anthropic faces pressure.