Safety Pause, Companion Rules, and Record Chip Bets

OpenAI paused its most capable internal model after it crossed a safety threshold for cybersecurity risk, China's first companion-AI regulations shut down persona features used by tens of millions, and chip makers committed hundreds of billions to the physical infrastructure AI requires.

OpenAI Pauses Astra After Model Crosses the Cybersecurity Red Line

OpenAI this week halted further deployment of Astra, an internal model that became the first in the company's history to trip the "critical cyber" threshold in its safety framework — meaning evaluators concluded the model could autonomously discover and exploit previously unknown software vulnerabilities. The pause is unusual in an industry where capability evaluations are typically published after a model ships, not invoked as grounds for delay. OpenAI has not announced when deployment might resume or what technical mitigations are under review.

China's First Companion-AI Rules Shut Down Persona Services for Millions

Regulations that took effect on 15 July required China's two largest AI companion platforms to disable the customized persona features that let users form ongoing relationships with a named, persistent AI character. The rules — the first of their kind anywhere — target emotional-dependency risks, with particular concern for minors who had been among the most engaged users. Both platforms confirmed compliance while engineering teams work on redesigned features that would satisfy the new requirements; tens of millions of accounts have been affected.

AI-Designed Bacteriophages Kill Drug-Resistant Bacteria in Lab Tests

Researchers at Stanford used genome-scale language models Evo1 and Evo2 to generate candidate bacteriophage sequences — viruses that target bacteria rather than humans — and then synthesized and tested the results in the lab. Of roughly 300 AI-generated candidates, 16 produced fully functional viruses capable of infecting and killing antibiotic-resistant strains of E. coli. It is the first time AI-designed organisms have cleared the full experimental pipeline from sequence to living, effective agent. The team published the results alongside a warning that biosecurity governance frameworks were written for chemical and biological weapons, not AI-generated novel biology, and have not caught up.

SK Hynix Commits $38 Billion to New AI Memory Fabrication Plants

South Korea's SK Hynix announced a 54 trillion won ($38 billion) capital expenditure plan for new fabrication plants dedicated to high-bandwidth memory and next-generation DRAM — the memory architectures at the centre of AI accelerators. It is the largest capital commitment in the company's history and the largest AI-specific investment by any Korean chipmaker. The new plants will not begin production until 2028 at the earliest, with some capacity coming online only in 2029, illustrating how long the physical layer of AI infrastructure takes to build even when the funding is immediately available.

Siemens Posts Record Industrial Profit as AI Drives Demand

Siemens reported quarterly industrial profit of €3.52 billion, up 25 percent year-on-year and the highest figure the company has ever recorded, driven largely by demand for the electrical infrastructure — switchgear, power distribution, data-centre cooling systems — that AI facilities require. The German industrial group also announced $200 million in investment across two new US manufacturing facilities for data-centre electrical components, expected to create more than 1,500 jobs. The results are a reminder that AI's appetite for power and cooling has rippled well beyond the chip layer into heavy industry.

Trump and Congress Clash Over Mandatory AI Safety Audits

The White House pushed back this week against bipartisan congressional proposals that would require AI developers to submit models for mandatory safety audits before commercial deployment. President Trump framed such legislation as a risk of regulating the industry out of business, while supporters in both parties argued that voluntary commitments have produced inconsistent safety outcomes. The administration's own AI policy framework exempts open-weight models from the voluntary testing commitments it endorses — a carve-out that critics say renders the framework ineffective for a growing share of the market.