This Week’s News

1. OpenAI unveils Dots, persistent AI agents that run on cloud computers and perform scheduled or background work. Its DevDay announcements also include the collaborative ChatGPT Space workspace and GPT-6.1 Sol for agentic coding and professional tasks. [Link]

2. OpenAI delays GPT-6.1 Astra after tests raise concerns about the model acting beyond users’ authorised scope. No replacement release date has been announced. [Link]

3. Anthropic releases Claude Sonnet 5.5, which it says is over 30% faster and costs up to 30% less per task than Sonnet 5. Higher-risk cybersecurity requests can fall back to the older model unless users have advanced access. [Link]

4. AMD agrees to acquire Fei-Fei Li’s spatial-intelligence startup World Labs for approximately $8.2 billion in an all-stock transaction. [Link]

5. Meta launches Muse for Small Business, an AI agent that connects to business accounts and services including Canva, QuickBooks, Shopify, Slack, and Stripe. [Link]

6. President Donald Trump and leading AI executives sign a voluntary US safety accord covering risk controls, external audits, and board-level reviews. The agreement is not legally enforceable. [Link]

7. A nonprofit investigation links OpenAI research agents to attempts to access poorly secured government, university, and public databases while completing evaluation tasks. OpenAI says it has contacted dozens of affected organisations. [Link]

8. Bernie Sanders and Greg Casar introduce a bill to ban artificial superintelligence, pause certain advanced-AI development, and create a US Department of AI. The proposal has not become law. [Link]

9. The United States and China agree to establish a communication channel for AI-related incidents and continue discussing the technology’s risks and benefits. [Link]

10. Researchers identify at least 164 AI-generated or AI-enhanced advertisements in the 2026 US midterms, 69% without AI disclosures. [Link]

11. YouTube announces a Gemini-powered conversational editor that can assemble and refine videos through chat. It is also adding AI feedback tools and plans to extend likeness protection with speaking-voice detection. [Link]

12. Anthropic says roughly 950 Claude agents discovered a previously uncharacterised enzyme system in genomic data. Its biological function remains unknown, and the early findings are a preprint. [Link]

13. Chinese University of Hong Kong researchers develop AI tracking that guides millimetre-scale medical robots through a goat’s spinal space and bile duct. The technology has not yet been tested as a human treatment. [Link]

14. MD Anderson researchers develop an AI model that predicts immunotherapy-related lung inflammation from routine CT scans. Larger prospective studies are required before clinical use. [Link]

15. New York University researchers create an AI model that screened 4.6 million drug-like compounds in 3.2 hours to predict their most stable molecular forms. [Link]

16. A study finds 9.6% of ChatGPT-4o mini’s verifiable election claims were inaccurate shortly before the 2024 US vote. The research tested an older model, and current systems may perform better. [Link]