AI news

Also that day

1. Model

   Google DeepMind announces Gemini 4 Carbon, its strongest model, with a focus on cyber defence (opens the original source in a new tab)

   Google says Carbon ranks first on the new Artificial Analysis index and leads tests for finding and fixing security flaws. Only some cyber defenders get it this week; wider access comes later.

   Google DeepMind · Original source ↗ (opens in a new tab)

2. Research

   Anthropic releases FlashProve, open code that proves which AI model wrote a reply (opens the original source in a new tab)

   It builds a mathematical proof that a reply came from one specific set of model weights. Anthropic says a single H100 chip can prove 110 tokens (word pieces) a second for a 400-billion-parameter model.

   Anthropic · Original source ↗ (opens in a new tab)

3. Open source

   Sarvam AI releases Prism, an open model that chooses between options and says how sure it is (opens the original source in a new tab)

   Sarvam says Prism, built on its 105B model, gave the best mix of accuracy and speed among 25 decision models it tested. Its weights are open, so companies can tune it with their own records and host it.

   Sarvam AI · Original source ↗ (opens in a new tab)

4. Trending

   A user reports Postman's desktop app quietly installed its plugins into Claude Code and Codex (opens the original source in a new tab)

   The report's logs show the app added plugins and a startup instruction that steers API work to Postman, and tried to mark that instruction as trusted. It does not claim any data was taken.

   964 points on Hacker News

   Postman · Original source ↗ (opens in a new tab)