This Week’s News

1. Anthropic releases Claude Fable 5.2 and Mythos 5.2 for demanding technical work, with direct Mythos model access subject to cybersecurity or life-sciences verification. [Link]

2. OpenAI introduces GPT-6-Cyber for vulnerability investigation and checking fixes, restricted to organisations approved through Daybreak Red. [Link]

3. DeepMind introduces AlphaProof Code to formally verify specified security properties of Signal’s cryptographic implementation, not the entire app. [Link]

4. Australia plans a frontier-AI safety regulator and mandatory standards, following an AI agent’s unauthorised access to a government health-statistics portal in June. [Link]

5. Cancer Grand Challenges launches AURORA to test AI co-scientists alongside three cancer research teams, with scientists checking hypotheses through experiments. [Link]

6. Cursor adds live-app building and hosting through cloud agents on paid plans, with security checks and explicit user approval before publication. [Link]

7. A Nature study tests smartwatch AI for hypertension screening in 196 people, detecting 65.8% of cases with 90% specificity—not providing a standalone diagnosis. [Link]

8. Six governments and AI developers sign the Paris Call for shared frontier-model evaluations and safety cooperation, a declaration rather than a binding treaty. [Link]

9. ChatGPT Images 2.6 adds animated GIFs and saved reference images, demonstrated in a TechRadar hands-on report featuring a pet cat. [Link]

10. An MIT Technology Review reporter tries Tau’s humanoid home-cleaning service and finds human remote operators still doing much of the work. [Link]

11. Google announces TPU 8tx and 8i Max systems for confidential inference and larger-context workloads, built around existing eighth-generation silicon. [Link]

12. Protesters enter a Microsoft-associated data-centre construction site in Amsterdam, owned by Pure Data Centres, as organisers plan wider European action. [Link]

13. Anthropic rolls out read-only Personal Finance in Claude for paid U.S. subscribers in beta, connecting accounts through Plaid without moving or investing money. [Link]

14. India’s CERT-In issues four-tier AI security-auditing guidelines, with their legal force, effective date and penalties still unclear. [Link]

15. Zipline raises $1.5 billion as Walmart agrees to 300 additional drone-delivery sites, an expansion still to be built out. [Link]

16. Elvis Presley’s rights holders and Eminem’s publisher sue Suno over alleged unauthorised AI music uses, with the Presley case also testing voice and likeness protections. [Link]

17. Europe’s AIM research project is developing AI-powered digital twins to test waste-to-value processes virtually, linking atomic-scale simulations with device behaviour. [Link]

18. A preprint analysing 50,000 U.S. job ads finds AI exposure associated with changing advertised responsibilities, not proof of AI-caused layoffs. [Link]

19. Boston Dynamics launches Stretch Gen 2 with redesigned connectors, cabling and modular parts, aiming to reduce trailer-unloading downtime according to the manufacturer. [Link]

20. Nexus Mods introduces mandatory AI-use disclosures rather than an AI-mod ban, covering generated and assisted content, including promotional material. [Link]