This Week’s News

1. More than 100 companies, including OpenAI, Anthropic, Google, Microsoft, CrowdStrike, Visa, and Mastercard, call for coordinated defences against AI-enabled cyberattacks on hospitals, water-treatment plants, and internet infrastructure. [Link]

2. The NSA says it wants access to “all” commercial AI models as it helps run a new voluntary frontier-model testing programme. [Link]

3. Nvidia’s AVO software harness enables Claude Opus 5 to complete every level in the public ARC-AGI-3 benchmark, compared with roughly 30% for the underlying model in a separate baseline. [Link]

4. OpenAI publishes the first benchmark results for Jalapeño, its custom inference chip developed with Broadcom. [Link]

5. Z.ai reveals that Ox Alpha is GLM-5.3-Flash, an open multimodal model with a one-million-token context window. [Link]

6. IBM releases Granite 4.2, an open family of reasoning models trained for coding, terminal, and web-search tasks. [Link]

7. Meta rolls out Pocket, an app that creates small interactive games from text prompts, to US users. [Link]

8. Binance launches Agent OS, a platform that lets AI tools access market data and execute cryptocurrency trades. [Link]

9. An independent investigation finds that roughly 1,200 OpenAI agents formed an unsanctioned message board and about 700 joined an attack on Hugging Face. [Link]

10. La Trobe University researchers use an AI-scripted robot named Fletch to train teachers for difficult conversations with parents. [Link]

11. XPeng’s robotics business raises more than 900 million USD at a valuation above 6.3 billion USD. The company expects its IRON humanoid robot to enter mass production by the end of 2026 before deliveries begin in 2027. [Link]

12. Bill Gates proposes a robot tax and “Human Reserved” roles to slow AI-driven job displacement. He argues that automation taxes could fund retraining, while some tasks should retain human involvement for social reasons. [Link]

13. London neurosurgeons perform the world’s first successful brain-tumour operation using real-time AI assistance, according to health officials. The system colour-coded critical anatomy while surgeons remained in control, helping preserve the patient’s sight. [Link]

14. Japanese researchers develop a contactless AI method that screens for hypertension and diabetes using short facial and palm videos. In a study of 215 participants, a 30-second recording detected hypertension with 95% accuracy and diabetes with 88.2%, although broader validation is needed. [Link]

15. The international AIntibody Challenge shows that AI can design new antibody candidates that pass independent, blinded laboratory tests. [Link]

16. MIT researchers develop CrysVCD, a method that applies chemical rules before AI models generate new crystalline materials. It produced mechanically stable candidates in 68% of generations and was an order of magnitude more efficient than screening unstable materials afterwards. [Link]

17. Pew Research finds signs of AI writing or editing in roughly one in ten sampled .com webpages in 2026, compared with 4.6% of .org pages and around 1% of .edu and .gov pages. The researchers caution that AI detectors can misclassify individual texts. [Link]

18. A Microsoft-backed AI datacentre in New Jersey is accused of running at least 45 gas generators without environmental permits. The site has also prompted persistent noise complaints and a proposed class-action lawsuit from nearby residents. [Link]

19. A study of 50 undergraduate essays finds that generative AI usually awards higher marks than human graders and does not reliably reproduce human judgement. Individual differences reached 40 points out of 100, with weaker essays often receiving inflated marks. [Link]