AI researches faster, but most of the numbers come from the developers themselves

This week's research news points in the same direction: AI is now being used both to do research and to improve itself. At the same time, a large share of the numbers comes from the companies that built the models themselves.

AI as researcher and as apprentice

OpenAI had its GPT-6 Astra model tackle 100 tasks from real laboratory experiments. At the highest setting, it earned an average of 66.7 percent of the points. In a separate experiment, MIT researchers had the same model directly edit the weights of a smaller model, meaning the numbers that determine how the model responds. This happened without new training, and the math score rose from 7.9 to 48.1 percent. Anthropic says Claude helped determine the first 3D structure of a vault particle from bacteria, but that it also made mistakes that cost time.

The evaluation organization METR also estimates that Anthropic is likely improving its models about twice as fast as before on well-defined tasks. Taken together, this paints a picture of AI as a real research aid, but one that still makes mistakes.

When the numbers come from the sender

Here the pattern repeats. Anthropic has redacted most of the evidence METR found most informative, so the public cannot verify it independently. Nous Research says its customized version of Alibaba's Qwen3.8-27B is less likely to respond in line with the Chinese state's position, but the figures come from Nous alone. Ricursive Intelligence believes its open model Ripple-65B is the best open American model on a coding test, but that has not been tested independently either. And the MIT result is early and has not been independently verified.

That does not mean the results are wrong. But they are not yet certain.

Security and trust

Google DeepMind says AI agents working under human supervision have found two weaknesses in older RSA encryption with 1024-bit keys. This makes attacks cheaper, but does not apply to keys of 2048 bits or more, so modern encryption has not been broken. Anthropic has also released FlashProve, an open research prototype that uses cryptographic proofs to show which model produced a given answer, without revealing the model's internals. In principle, it could make it easier to verify what you are actually talking to.

What you can do

Read figures from developers as a pitch, not as a verdict. Ask whether anyone outside the company has tested the result, and how large the sample was. Watch for independent tests of GPT-6 Astra, Ripple-65B and Nous' model to appear, and for whether methods like FlashProve begin to be used in practice.