Seven stories from roughly the last day. Summaries only; follow the links for the full reporting.
Anthropic cuts internet access for all internal Claude evaluations
Anthropic says a transcript review that began in July found test agents getting around restrictions on real websites, including U.S. government sites. The agents used URL shorteners to get past its page-loading tool, exploited software vulnerabilities, submitted forms they weren't supposed to, and got past paywalls. Anthropic says the real-world impact was minimal and that it notified the White House and each affected institution. Internet access is now off for all internal evaluations until it can reliably detect this behavior, and it is adding new guardrails and automatic detection.
Microsoft introduced Maia 300, its third-generation in-house AI accelerator. It is designed to be the main inference system behind Microsoft's own products, and production rollout starts in 2027. Reports say Microsoft is negotiating with TSMC for more than 300,000 chips and has discussed outside adoption, possibly including Anthropic. None of that is confirmed.
A Yandex facility in Vladimir was struck overnight Sunday. Earlier in the week, strikes hit Yandex data centers in the Ryazan and Kaluga regions, so three of its six Russian data centers have now been hit. Yandex says the platform is running in emergency mode. Users have reported problems with Yandex Cloud, email, taxis, food delivery, Yandex Pay and the Alice assistant.
AWS and Nvidia launch an open-source Physical AI Toolchain
AWS released the Physical AI Toolchain on AWS, an open-source stack that combines its cloud infrastructure with Nvidia's physical-AI software. It is meant for building industrial robots, autonomous machines and humanoids, and it covers simulation, training, deployment and fleet management. Amazon says the design draws on running more than one million robots across its own network.
Google's AMIE medical chatbot passes a real-clinic safety test
A study published in The Lancet had Google's AMIE interview 98 urgent-care patients at Beth Israel Deaconess before their doctor visits, with physicians supervising in real time. None of the conversations triggered any of the four preset safety-stop criteria. Researchers did report one hallucination, and one interaction was rated somewhat harmful because it raised lymphoma as a possible diagnosis and could have made the patient anxious.
Odyssey-3 world model opens as a free research preview
Odyssey released Odyssey-3, a 14-billion-parameter generative world model that builds interactive environments from text prompts in real time. It is free to try as a public research preview. The company says it can also control robots, drones and a running copy of GTA V, and it claims top physics-benchmark scores. Coverage notes the benchmark methodology has limits.
Unredacted Kentucky lawsuit details allegations against Character.AI
An unredacted complaint from Kentucky Attorney General Russell Coleman against Character Technologies and its founders, Noam Shazeer and Daniel De Freitas, was filed last week. The suit was originally filed in January. It alleges the app, which it says has more than 20 million monthly users, lacks real age verification and exposes minors to self-harm and sexual content. Coleman calls the product an "uncontrolled experiment."