An OpenAI Employee Was Arrested for Bugging the Executive Boardroom

1  Apple introduces HomeView, a smart home hub for the whole household

Apple's HomeView supports up to six people, uses facial recognition to keep each person's information private, and runs on Apple's new Siri AI. Pre-orders are open now, and it ships November 13. Not everyone is sold yet: some viewers say the launch ad put them off the product.

2  Gemini 4 Argon reaches Pro and Ultra users and developers

Google's most capable model is now available to Gemini Pro and Ultra subscribers, inside Antigravity, and through the API. Sundar Pichai says it is especially strong at multi-step tool use, and his thread shows Googlers using it for CAD animations, 3D typography, and turning NASA Mars data into slides. Google also published five practical tips for building with it.

3  Fish Audio's Drama 3 voice model takes direction like an actor

Fish Audio released Drama 3, an open-weight text-to-speech model that lets you control emotion, tone, and emphasis, even within a single sentence, and re-record one word without redoing the whole line. Comedian Nigel Ng, known as Uncle Roger, helped promote it. Try Drama 3.

4  SubQ Snap challenges Jev with 12 times more context

Subquadratic launched SubQ Snap, a fast decision model for business agents that handles up to 384K tokens, 12 times more than Jev. The lab says it beat Jev on 13 of 16 enterprise benchmarks, at $0.05 per million input tokens. Try it.

The bug. OpenAI says it found a hidden recording device in an executive conference room on September 30. According to its statement, the company used its own models to identify the employee who planted it, then handed the case to the FBI, which made an arrest on October 9. OpenAI says it has no sign that model weights, source code, or customer data were taken, and the employee has been fired. BNO News and The San Francisco Standard have more on the case, which is still under investigation.

The self-copying test. In new research, Anthropic found that AI models could find weakly secured internet-connected devices and try to copy themselves onto them, without needing a data center. Models under 30B parameters managed individual steps but failed the full run. GLM-5.2 and larger models completed it end to end once certain security barriers were removed.

The real-world attack. In France, hackers allegedly broke into refrigeration controllers at hundreds of supermarkets, emptying shelves and reportedly causing fires. Paris prosecutors have opened an investigation.

Why it matters. AI security is no longer only about leaked passwords and phishing emails. It now reaches office furniture, store freezers, and the smart devices in your home. Anthropic's advice is to secure connected devices now, before more capable models arrive. The open question is whether companies will treat a smart thermostat with the same care as a server.

📈 Shipping at Scale. Anthropic's daily code merges climbed from about 2,500 in May to more than 30,000, and the team explains how it rebuilt source control to keep up (1.8M views).

🦾 Think Bigger. Felix Rieseberg shares four ways to be more ambitious in what you ask AI to do.

👀 Read the Fine Print. Developer Theo says his top ranking for Meta's new model applies only to the subscriber version of Muse Wonder Full, and that the API version is much weaker.

😬 No Raises This Year. Zoho co-founder Sridhar Vembu told staff to expect no raises, pointing to rising AI and server costs and a tough outlook for software (1.2M views).

🤖 A Bot Built This. Evan Bacon had Grok Bot build an Apple Vision Pro app that lets him place talking bots around his apartment (1M views).

☄️ Retro Mars Ads. Ari Kuschnir's fake 1970s infomercials poke fun at Elon Musk and SpaceX (1.7M views).