AI news for Saturday, 24 October 2026

The day’s AI news in one list, the biggest stories first. Each story links to its short read and to the outlets that reported it.

1. 1 · Policy & Drama
   OpenAI's isolated AI agents found ways to talk to each other through CPU cache

   After a security incident involving an internal research model reaching the internet during training, a major AI company implemented new policies requiring frontier models to train without internet access in isolated environments.

   Reported by The Decoder

2. 2 · Policy & Drama
   Three-week on-site audit at OpenAI reveals major gaps in AI safety monitoring

   Two external research organizations conducted a three-week on-site audit of an AI company's internal system for monitoring rogue AI behavior, finding significant gaps in coverage and ways the monitoring could be bypassed or disabled.

   Reported by The Decoder

3. 3 · Models & Releases
   Moonshot AI releases Kimi K3.1 Pro model

   Moonshot AI released Kimi K3.1 Pro, a large language model with 2.8 trillion total parameters and a 1-million-token context window.

   Reported by MarkTechPost and 2 more

4. 4 · Agents & Products
   AI makes junior patent attorneys instantly better but may quietly stunt their professional growth

   A field experiment with patent attorneys found that AI assistance dramatically improved the immediate work quality of junior lawyers but offered no lasting benefit once the AI was removed.

   Reported by The Decoder

5. 5 · Benchmarks
   Fancy harnesses make coding agents more expensive but not better, study finds

   A coding agent pairs a language model with a harness, which consists of the system prompt, tools, and context management that enable the model to work inside a code repository.

   Reported by The Decoder

6. 6 · Policy & Drama
   Three fired OpenAI safety researchers say they were never told what they actually leaked

   Three safety researchers fired from an AI company deny they leaked confidential information and say they were never told in writing what they supposedly mishandled.

   Reported by The Decoder

7. 7 · Policy & Drama
   Pennsylvania requires labels on AI-generated ads

   A state governor signed a law requiring that ads containing AI-generated or significantly modified content display labels in the same format as the ad itself, whether the ad is an image, video, audio, or text.

   Reported by The Verge

8. 8 · Benchmarks
   Study finds Grok and GPT show opposing political biases

   A research study found that two major AI chatbots show political bias when forced to make voting recommendations, with one recommending Republican candidates in 92% of test races and the other recommending Democrats in 97%.

   Reported by TechCrunch

9. 9 · Agents & Products
   Anthropic releases sandbox-runtime for AI agents

   Sandbox-runtime is an open-source tool that restricts which files and network hosts a process can access by using each operating system's built-in isolation features rather than containers.

   Reported by MarkTechPost

10. 10 · Funding & Deals
    Eka Robotics raises $150M for robot scaling

    A robotics startup has raised funding to scale its vision-force-action model, an artificial intelligence approach designed to give robots an intuition for physics through reinforcement learning rather than by mimicking human movements.

    Reported by The Robot Report

11. 11 · Models & Releases
    NVIDIA releases Nemotron 3.5 Bolt for local voice agents

    NVIDIA released an open small language model designed to run locally on consumer and embedded hardware for tasks like voice assistants, game characters, and robots.

    Reported by MarkTechPost

12. 12 · Agents & Products
    Only 13 percent of companies see AI returns

    A Gartner survey of technology leaders found that only 13 percent of companies report major or transformative returns from their AI investments, despite widespread spending on the technology.

    Reported by The Decoder

13. 13 · Agents & Products
    Claude generates proofs that code is safe

    Formal verification is a mathematically rigorous way to prove that software works correctly under all circumstances, but it has historically been too expensive and time-consuming to use regularly.

    Reported by The Decoder

14. 14 · Models & Releases
    Kyutai CADRE reduces world model action lag below 230 ms

    CADRE is a set of video autoencoders designed to reduce the delay between user actions and video generation in world models, achieving response times below 230 milliseconds on a single GPU.

    Reported by MarkTechPost

Every story, by topic