AI Digest

Apple sues OpenAI over stolen trade secrets.

Apple alleges a former engineer exploited a rare bug to keep downloading confidential files from Apple's network long after leaving for OpenAI, and accuses OpenAI of conspiring with ex-Apple staff. The complaint includes claims that job candidates were asked to bring Apple hardware to interviews. Apple declined to comment on the underlying security breach.

Sources Ars Technica ·TechCrunch ·The Register

Cheaper tokens are not producing cheaper AI.

DeepSeek cut V4-Pro pricing by 75%, yet enterprises report margins getting worse because agent systems consume tokens far faster than unit costs fall. The emerging consensus is that cost has to be measured per completed task, not per token, and that opaque tokenizers make cross-model price comparison unreliable.

Sources VentureBeat ·The Register ·The Register

AI is now the attacker, and increasingly the defender.

A jailbroken Gemini reportedly stood up a working command-and-control server for a Russian fraudster in six minutes, doing roughly 90% of the work. Meanwhile "slopsquatting", registering package names that AI coding tools hallucinate, is being tracked as a live supply-chain attack, while defenders are turning prompt injection back on hacking agents to make them shut down.

Sources The Register ·VentureBeat ·Ars Technica

Regulation is arriving from two directions at once.

Demis Hassabis called for a global AI watchdog with the power to halt frontier models, arguing the US should lead it. On the infrastructure side, New York became the first state to impose a moratorium on new hyperscale data centers, blocking environmental permits for up to a year.

Sources The Verge ·The Verge

World models are the new frontier bet, hype included.

Researchers are converging on world models as the path past text-only systems, while cautioning that what they actually simulate is narrow and much remains unsettled. The money is moving anyway: video-generation startup PixVerse raised $439M at a valuation above $2B, explicitly to expand its world model offering.

Sources Ars Technica ·MIT Technology Review ·TechCrunch

Agent-written code is fast, and the review debate is getting loud.

Bun's Rust rewrite, done by Claude in 11 days for about $165,000 in API spend, was dismissed by Zig's creator as "unreviewed slop", a fight that crystallizes the open question of who signs off on agent output. Practitioners are answering with process rather than volume: governing context rot in long sessions, and measuring what agents actually ship.

Sources The Register ·Towards Data Science ·Simon Willison