- Anthropic will close down chats for «abusive or cruel behavior» toward Claude [The Verge]
- Frontier labs donate $2.4B to Trump’s Genesis Mission to advance federal science agencies [Politico]
- Arena.ai releases Arena Alignment Index benchmark, finds misalignment is rare, but has serious consequence [Arena/X]
- Anthropic launches «Cyber Mission» to secure critical infrastructure and open-surce software [Anthropic]
- OpenAI tell investors it’s $20B behind revenue target. AI stocks fall on bloody market day [CNBC]
- Google launches SynthID Detector, to check for watermarks from Google, OpenAI, Nvidia and «soon» Apple [Google]
- Google intros Playground, a web interface to create AI games from prompts, in the US [Google]
- Anthropic releases Claude Haiku 5.5, improving on 4.5 and 75% cheaper for high-volume tasks [Anthropic]
- Meta, DeepMind and Isomorphic Labs to standardize govt biological data for AI training [Reuters]
- Microsoft launches $2,599 Surface Laptop Ultra, powered by Nvidia’s RTX Spark [Microsoft]
- Meta, Sierra, Walmart and others develop open Personal Agent Protocol to govern interactions with shops, business [Sierra]
- Mistral previews 1T Mistral Large 4, Le Chonk, that «pushes the frontier of open-weights» [Mistral]
- Cheap AI tools are flooding black markets and dark web, report finds [Axios]
- Italian Prime Minister Meloni applies to trademark her voice to combat deepfakes [AP]
- ChatGPT not only mimicks New Yorker cartoons, but adds the artists’ signatures [Niemanlab]
- Google DeepMind claims to design new-to-nature enzymes from scratch [Pushmeet Kohli/X]
- Wikimedia says OpenAI agents edited some wikis and failed to compromise note-taking tool [Engadget]
- Reflection AI unveils Beam, an open model approaching Qwen3.8-Max on coding and agentic tasks [Reuters]
- Anthropic refers Florida woman to police for Claude threat to «shoot up» Sheriff’s office [Tom’s Hardware]
- OpenAI says default speed across GPT-6 Astra and 6.1 Sol are now 50% faster [Tibo/X]