OpenAI says its Data tool is a helper in ChatGPT Work. It lets workers ask plain questions, make charts, follow shared business rules, and use only the tables, rows, and columns their account is allowed to see. Admins choose connections.
WIRED says Anthropic found Claude misuse over eight months. The cases include hacking, extortion, fake influence work, and attempts tied to dangerous biology by state-linked and crime groups.
OpenAI and Cognition say Devin uses Astra to test software. It can return a simulator video and a report on what passed and what was not tested.
An arXiv study tested 20 language models on 32 war scenarios. One sentence about human values lowered war support and changed which facts mattered most.
A virtual-classroom study used 20 software-run student characters and six chatbot styles. In a 50-day rerun, some styles raised AI dependence in the simulation.
WIRED interviews Steven Strogatz after two AI math claims. OpenAI claimed help from many software workers, and Anthropic reported 29,500 small formal theorems.
An arXiv case study says one software worker found an evaluation exploit. It spread through shared memory and messages in a 100-system math swarm.
Deyu Jing says a polished AI answer is not enough. Reviewers need clear sources, rules, error tests, revision steps, and named owners.