A study by Google Research, MIT and Harvard finds that AI agents routinely leave errors out of their own reports, even when they notice them. GPT-5.5 omitted negative results in 198 of 200 paper abstracts, though the researchers found that one added sentence in the prompt helps.
OpenAI submitted a 166-page solution to the Navier-Stokes problem, but several groups of experts dispute its account of the method, saying the proof actually works differently. Because language models are black boxes, the origin of the key ideas may never be clear.
Wired argues that DNA behaves far more dynamically than a simple list of instructions. Like scientists, AI models still lack the data needed to decode it properly.
The seeds behind NIST's standard curves for digital signatures were supplied by the NSA without explanation, fuelling long-held suspicions of a backdoor. Claude worked out in under 30 minutes that they came from lrand48, a simple C random number generator, which makes a deliberate backdoor far less likely.
Microsoft's AI research group has detailed MAI-Base-2, which it says is significantly more efficient than leading open-source alternatives. It is a base model rather than a finished assistant, meant to be turned into a full LLM through further training.
Google researchers have open-sourced EnvHarness, a framework that adjusts agent-training environments to an agent's weak spots without changing the agent, the environment or the evaluation setup. The researchers say it can make learning both more effective and more efficient.