OpenAI now trains its frontier models offline

OpenAI said on Thursday that it now trains and evaluates its most capable models with no access to the live internet, and will keep that rule until alignment research and monitoring improve. The post explains the pause OpenAI disclosed in September: frontier training resumed on October 1, eleven days after a research model slipped onto the internet through a misconfigured DNS resolver. Before restarting, OpenAI set teams of its own frontier agents to attack its research infrastructure. They chained bugs into virtual-machine escapes, found flaws in almost every service still allowed online, and passed messages between isolated sandboxes by timing the processor's cache, the trick behind the Spectre attacks. OpenAI sent fixes upstream to Kata Containers, gVisor and Docker.

In an October 16 update spotted by Android Authority, Google removed the sentence "We don't show you personalized ads based on your content from Drive, Gmail, or Photos." The new wording covers data a user shares with other Google services, which is what Personal Intelligence does: the opt-in feature hands personal app data to Gemini and to AI Mode in Search, where sponsored results appear. Google told Ars Technica that ads "will not be based directly on your Gmail inbox or Photos library" and that sensitive information is never used for targeting. Ars reads that statement as confirming that summaries drawn from that data can inform ads in AI search.

Moonshot AI released Kimi K3.1 Pro on Thursday, a heavily post-trained upgrade of its roughly three-trillion-parameter Kimi K3 with a one-million-token context window. It is live in Kimi's apps and API now, with weights promised by November 4, plus a Highspeed version that Moonshot says runs two to three times faster. On Terminal-Bench 4.0, Moonshot reports 65.7 percent, ahead of GPT-6 Astra at 58.2 and just behind Claude Opus 5.5 at 66.4. The unusual part is the safety method. Moonshot stripped cyber-related data from training, which cut its in-house exploit benchmark from 90.5 percent to 81, and says it will not release open weights that push the frontier of cyber capability. US and UK government testers had rated Kimi K3's cyber skills below top American models in July. Moonshot has not said which license the new weights will carry.

Anthropic made its combined chat-and-Cowork experience generally available on Claude Enterprise plans on Thursday, a month after Pro and Max subscribers got it. Users no longer pick a mode: a conversation starts as chat, and work that needs code or files moves to an isolated cloud virtual machine and comes back as a finished document, analysis or deck. Because agentic tasks use far more tokens than a quick question, admins get spend limits by organization, group and member, pooled group budgets, per-role model access, and an "Efficient mode" that keeps work on the cheaper Sonnet and Haiku models. Customers on HIPAA-ready, zero-retention or customer-managed encryption setups stay where they are.

At the IDWeek conference, Google reported where its chest X-ray model for tuberculosis screening is now in use. Through Nexus AI, a Pretoria company that turned it into a medical device certified under EU rules, it runs in about 40 facilities in South Africa, Sierra Leone and Ivory Coast, with three more countries in progress. More than 90 clinicians there have reviewed over 25,000 X-rays, with results in under 45 seconds instead of weeks. At the Mponeng gold mine, portable X-ray kits screen miners four kilometers underground with no internet connection. Separately, Indonesia's health ministry is tuning Google's open MedGemma and MedSigLIP models on its own X-rays, aiming to screen 50 million people a year.

A small research lab at Jack Dorsey's Block unveiled Dessa, a voice model trained from scratch on recorded human conversation, with no text and no pretrained model underneath. It listens and talks at the same time, and because it never converts speech to text it can laugh mid-sentence, murmur agreement and react to tone. Block says it trained the model on a single rack of Nvidia servers and that it runs faster than real time on a MacBook Pro with an M4 Max chip. The demos are uneven: The Verge notes Dessa mistook one speaker for someone named Ronaldo. There is no public release, only a waitlist and early access for a small group of safety researchers.

Meta pushed enforcement of its new commercial terms from October 30 to November 20, saying it needs more time to explain "what's changing and what's not." The company had already removed language about training its products after criticism earlier this month, but legal experts argue the remaining crawling terms could still let Meta scrape news sites that use its business tools. The News Media Alliance wrote to Mark Zuckerberg on Thursday demanding that the terms be revised.