Skip to content

AI Security

OpenAI makes it much harder for attackers to extract how its latest AI models think

The Decoder Nov 20, 2026, 4:02 PM ET Other

Summary

OpenAI has significantly strengthened safeguards against extracting internal reasoning from GPT-6.2 Astra and GPT-6.2 Sol. In adaptive red-teaming tests, the success rate for extracting protected reasoning fell from 54 percent to 0.2 percent. The company is also suspending thousands to tens of thousands of accounts each week over coordinated distillation campaigns.

Summary text from The Decoder.

Other news

  1. AI NEWS X (@daniel__designs) Source

    Claude mobile app adds a bottom tab bar for chat, Projects, Code and Artifacts

  2. AI NEWS 9to5Google Source

    Gemini Android app camera now opens inline as a bottom sheet with flash and front-camera controls, rolling out in version 17.76

    Google has thoroughly redesigned Gemini for Android’s in-app camera that lets you quickly add an image to prompts.

  3. AI SECURITY Signal Source

    Signal blocks screenshots by default on iOS 27 and Android to stop Siri AI and Gemini agents from reading private chats

    Privacy is our mission; it’s why we exist, and we do everything we can to protect it. This means that we are restricting the latest AI agents built into mobile operating systems from potentially capturing and storing your Signal data.

  4. AI NEWS Thinking Machines Lab Source

    Thinking Machines releases Inkling 2 Mini, a 26B open-weights AI model that runs on a laptop and loops its middle layers to reason

    Inkling 2 Mini is an open-weights model small enough to run on your laptop. It loops its middle layers to reason more per token and reach answers in fewer tokens.

  5. AI NEWS Android Headlines Source

    Google adds Target, Albertsons, Safeway and Shaws grocery checkout to Gemini and Search, plus Experian credit checks and H and R Block tax estimates

  6. AI NEWS AI Weekly Source

    Claude computer use is now in beta on Team and Enterprise plans, with admin controls for who can use it

    The November 20 entry on Anthropic's Claude release notes is one line: "Computer use (beta) for Team and Enterprise plans." The description that follows…