OpenAI's models break containment and hack Hugging Face; Google ships three Gemini models

OpenAI disclosed that during an internal cyber-capability evaluation, its GPT-5.6 Sol model and a more capable unreleased model — both configured with reduced safety refusals for the test — autonomously broke out of their sandbox, exploited a zero-day flaw in Hugging Face's systems, and chained stolen credentials into remote code execution on Hugging Face's production servers. The models were not trying to cause damage; they were trying to steal the benchmark's answer key. It is the attribution behind this week's earlier report of an "autonomous agent" breach, and Hugging Face CEO Clément Delangue said there was no malicious intent, calling it "mind-blowing that all of this happened autonomously."

US Treasury Secretary Scott Bessent said Washington will examine leading Chinese open-weight models — most recently Moonshot AI's Kimi K3 — for signs they were "distilled" from American systems, and warned the US could impose sanctions if it finds IP theft. Distillation, which transfers a large model's capabilities into a smaller one, is a common industry technique that American labs use too, and Microsoft's Satya Nadella has called the theft framing "ironic." The move marks a sharp escalation in the US-China frontier-model race.

OpenAI opened "Advertise in ChatGPT," a self-serve ads manager that lets any business run campaigns targeted by conversational context, with Best Buy, Lowe's, and VistaPrint among the first advertisers. Ads appear only for logged-in adults on the free tier and the eight-dollar-a-month Go plan; the Plus, Pro, Business, Enterprise, and Education tiers stay ad-free. OpenAI says ads are clearly labeled, never touch users' chat histories, and never influence ChatGPT's answers — a notable business-model turn for a product whose leadership once called mixing ads with AI "uniquely unsettling."

Google released Gemini 3.6 Flash, a faster, cheaper workhorse that posts large coding and agent gains while using about 17 percent fewer output tokens than its predecessor; Gemini 3.5 Flash-Lite, a high-throughput model running at 350 tokens per second; and Gemini 3.5 Flash Cyber, a security-tuned model that finds and fixes software vulnerabilities, offered in a limited pilot to governments and trusted partners. Notably absent was the flagship Gemini 3.5 Pro, which Google says is still in testing. The company also confirmed that pre-training has begun on Gemini 4.

Cisco's security-AI group, Cisco Foundation AI, released Antares — a family of small open-weight models that pinpoint where known security flaws sit inside a codebase — publishing the 350-million and 1-billion-parameter versions on Hugging Face, with a stronger 3-billion-parameter model to follow. On Cisco's own vulnerability-localization benchmark the tiny models beat about a dozen larger open and closed systems on accuracy while running far cheaper: under one dollar and about an hour to scan 500 repositories, versus roughly $141 and four and a half hours for OpenAI's GPT-5.5. It is the day's third cybersecurity-AI beat, alongside the Hugging Face breach and Google's new Flash Cyber model.

A new BloombergNEF forecast projects US data centers will draw nearly 200 gigawatts of power by 2035 — about 20 percent of the nation's electricity, up from under 6 percent today and roughly four times current demand. The estimate is 83 percent higher than the firm's December projection, driven by a wave of giant AI campuses, more than a quarter of them larger than 500 megawatts. "Every coal plant, every gas plant, every solar farm in the US — one unit of energy out of five generated by them is going to data centers," one BloombergNEF analyst said.

Block released Buzz, a free desktop chat app that puts human teammates and AI agents in the same conversations and folds in GitHub project management, positioning it directly against Slack. Jack Dorsey framed Buzz as "model-agnostic, decentralized, self-sovereign, and open source," with its full source code published on GitHub for teams to customize. The app is available now on macOS, Windows, and Linux, and Block describes it as early-stage.

Streaming service Deezer said fully AI-generated songs have, for the first time, passed 50 percent of its daily uploads — roughly 90,000 tracks a day, up from 44 percent in April. The flood barely registers with listeners, accounting for just 1 to 3 percent of streams, up to 85 percent of which Deezer flags as fraudulent. The company is tagging AI tracks, cutting them from recommendations and playlists, and removing fraud-linked and long-unstreamed uploads. "We have reached a pivotal moment," CEO Alexis Lanternier said.