Agents strain AI safety tests.
Project No. 16AZW016 is real, 350,000 yuan is the standard funding amount, the project was rated excellent upon completion, and the widely circulated claim that "Mo Yan angrily lashed out" is false. I also scraped 371 Zhihu answers for content analysis: in what looked like overwhelmingly one-sided public opinion, 96% of the negative volume actually came from just two viral posts.
The author of Oh My Pi used an external tool to bypass the hidden reasoning protections of GPT-5.6 and Fable 5, exposing not only the thought process but also API credentials
A wider rollout could reshape expectations for consumer AI tools.
NextSlide staff are now working on ChatGPT.
AI helps solve a 25-year math problem.
AI security is moving from hype to practical validation.
AI adoption is running into hard cost math.
A talk outlines a governable defense model for AI agents.
False positives are reshaping how people prove their work.
