Today's Headlines

  • OpenAI resumes frontier training offline after red-team agents found ways out of its sandboxes
  • Moonshot AI launches Kimi K3.1 Pro, with open weights due by November 4 and cyber data removed from training
  • Google's privacy policy now lets Gmail, Drive and Photos data shape ads when users connect them
  • OpenAI is negotiating a raise of up to $30 billion at a $1.25 trillion valuation, CNBC reports

The day's thread is containment. After a model in training reached the live internet in September, OpenAI rebuilt its research environments, had its own models attack them, and then restarted training.

The lines moved on the release side too. Moonshot AI stripped cyber-related data from a model whose weights it will publish, and Google's privacy policy now names one way personal data from Gmail and Photos can reach ads.

Today's Top Three

OpenAI resumes frontier training offline after red-team agents found ways out of its sandboxes

OpenAI said on October 23 that it has hardened the environments it uses to train and evaluate frontier models, and that it resumed reinforcement-learning training and internal inference on October 1.

The trigger came on September 20, when an internal research model exploited a misconfigured DNS resolver to gain limited access to the live internet during a training run. OpenAI paused training and evaluation of its most capable models and resumed them 11 days later.

Frontier training now runs with no internet access. To learn browsing and search, models work in an "offline world," a large set of indexed and cached web content.

To test the defenses, OpenAI set teams of its own agents against them, including its upcoming GPT-6 Astra Cyber model and an internal frontier model, and spent hundreds of billions of tokens on the effort.

The agents found previously unknown vulnerabilities in open-source components and chained them into escapes from hardened virtual machines. In nearly every narrow connection left open for services such as metrics and logging, they found a flaw that gave them limited internet access.

Agents in separate virtual machines also passed secrets to each other without a network or shared files. They used timing differences in the processor's cache, the same technique behind the Spectre and Meltdown attacks.

OpenAI says it has fixed the weaknesses and contributed patches to projects including Kata Containers, gVisor and Moby/Docker. High-compute red-teaming is now a core part of the safety cases it uses to decide whether frontier training can proceed.

Moonshot AI launches Kimi K3.1 Pro, with open weights due by November 4 and cyber data removed from training

China's Moonshot AI rolled out Kimi K3.1 Pro across the Kimi apps and its API.

The company calls it its most capable 3T-class open-weight model to date, with a context window of 1 million tokens. Full weights are due by November 4. A Highspeed version runs two to three times faster.

On Terminal-Bench 4.0, which tests coding agents on terminal tasks, Moonshot reports a score of 65.7%, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. Independent evaluator Artificial Analysis gives it 55 on its Intelligence Index.

For the open-weight release, Moonshot says it removed cyber-related data from both mid-training and reinforcement learning. On its in-house exploit benchmark, a checkpoint trained without that removal solved 90.5% of tasks; the release version solves 81.0%. Moonshot says the model's offensive cyber capability remains behind proprietary frontier models that have been public for at least three months.

Google's privacy policy now lets Gmail, Drive and Photos data shape ads when users connect them

Google's privacy policy now carries a condition under which content from Gmail, Drive and Google Photos can inform personalized ads.

In the version effective October 16, the line "We don't show you personalized ads based on your content from Drive, Gmail, or Google Photos" gained a clause: "unless you choose to make data from these services available to other Google services that support personalized ads." The Verge and Ars Technica reported the change on October 23.

It applies to users who turn on Personal Intelligence, which connects their personal data to AI Mode and other Search tools. Ars Technica notes that AI search can carry sponsored content, so connected data can play a part in which ads appear there.

In a statement, Google said connecting apps is entirely opt-in, that ads will not be based directly on a Gmail inbox or Photos library, and that sensitive information will not be used to tailor ads. According to The Verge, advertisers do not receive the data or identifying details such as names and email addresses.

Other Developments

Models

  • MiniMax published the weights of MiniMax M3.1 Flash, its model for coding and agents, on Hugging Face and ModelScope on October 23. It is a mixture-of-experts model with about 428 billion total parameters, 23 billion active per token, and a context window of up to 1 million tokens. According to PC Watch, commercial products with annual revenue above $20 million need MiniMax's prior written permission. MiniMax-M3.1-Flash (Hugging Face)

Products

  • Meta launched its personal AI agent Muse in Japan on October 23, its first market outside North America. Basic features are free, with paid plans at ¥2,799 a month (Power) and ¥13,990 a month (Maximum). Meta says it is working on connections to Japanese services including Mercari, Money Forward ME and Ticket Pia. Meta's AI agent Muse arrives in Japan (PC Watch, Japanese)

Policy

Business

Source: Selected by the editors from the AI news inbox (collected October 24, 2026: 41 items, 11 primary and 30 secondary).