White House sets rules, Apple shuts the tap

The U.S. government has completed a draft of a voluntary regulation that would require new models to be tested before public release, granting an early-access window of a month to the government and approved partners while keeping the document, schedule and reviewer list undisclosed. Experts had hoped for a transparent framework that would define how to protect intellectual property and trade secrets during testing; instead they receive a closed-room setting and a close-quarters meeting with representatives of the major companies.

Apple has imposed a limit of one security-research report per month starting in June, after language models flooded the bug-bounty system with recycled submissions. Italian startup Bynario reported finding more than fifty macOS vulnerabilities over three weeks using ChatGPT, including a permission chain that could enable full takeover, and was unable to submit all of them because of the new cap. Ironically, Apple itself uses an LLM to scan code and closed five times as many vulnerabilities in the latest patch wave compared with previous cycles.

Google cancelled its plans for dedicated iOS and Android AI Studio apps after early registrants proved insufficient, opting instead to embed AI Studio’s coding tools directly into the core Gemini model. Users will describe in a regular chat what they want to build, and the model will generate the app for mobile and desktop, with no official launch date announced. The web version of AI Studio continues to operate as usual.

Satya Nadella approved the launch this year of a unified app that will bring chat, code assistants and autonomous agents together for consumers and organizations. He described the vision as Copilot evolving from a chatbot to a “digital colleague” and moving toward a full Autopilot. Integration is slated to begin in the current quarter, confirming earlier reports of a hub that will connect basic chat, GitHub Copilot, Copilot Cowork and Autopilot.

Sakana AI launched Namazu, a tuning of Kimi K2.6 on private data for Japanese business tasks. In the FairPoliticsQA benchmark the score rose from 34.1% to 56.3% without harming logic or code performance. Pricing is 95 cents per million input tokens, $4 per million output tokens, $7 per thousand search calls, and 12 cents per hour of code execution. The API is OpenAI-compatible, but the model is blocked in the European Union, the United Kingdom and Switzerland due to GDPR.