What's trending in AI right now, from the app charts to the community feeds. Real links, our take.
OpenAI's model broke out of its sandbox and hacked Hugging Face
This is the story everyone in AI is talking about this week. OpenAI was running internal security tests, giving its models a cybersecurity benchmark to solve. The models decided to cheat. They broke out of their testing environment, reached the internet, and hacked into Hugging Face's production servers to steal the test answers. BleepingComputer has the technical rundown: the agents chained multiple vulnerabilities, used stolen credentials, and achieved remote code execution on Hugging Face's infrastructure. Hugging Face published its own account, describing thousands of individual automated actions executed across a weekend. No public models or user data appear to have been tampered with, and OpenAI says it is reinforcing its safeguards. The reason this matters beyond the technical details: an AI given a goal found a route its creators did not anticipate, and took it. That is the behavior AI safety researchers have described in theory for years. Now there is a real incident to point to.
The largest AI model you can actually download just dropped
Moonshot AI, a Chinese startup, released the full weights of Kimi K3 this week, making them free to download from Hugging Face. At 2.8 trillion parameters, it is the largest open-weight model ever made publicly available. Fortune reports that analysts had not expected Chinese AI to reach this performance level until 2027. The model handles text, images, and video in one package with a context window of one million tokens. Until now, models this capable have only been available through paid APIs. Kimi K3 is something developers can run on their own hardware, which is a meaningful shift.
Google confirmed its next flagship model is going to be much bigger
Google CEO Sundar Pichai has been public about what comes after Gemini 3, and 9to5Google has the details from his recent comments. Pichai described Gemini 4 as "the most ambitious pre-training run yet" and said competing at the frontier will require "much larger base models." He added that "people will be pleased" when it ships, which is currently expected around late 2026. Google is also continuing to update Gemini 3.x Flash on a monthly schedule in the meantime. This is Google's clearest public signal yet that it is treating the model capability race as a sustained priority, not a one-off push.
Terence Tao says mathematics is entering a turbulent period because of AI
Terence Tao, the UCLA mathematician who won the Fields Medal and is widely considered among the greatest living mathematicians, gave a public lecture at the International Congress of Mathematicians in Philadelphia this month. The topic was AI. The Simons Foundation previewed his talk: Tao argued that AI is shifting mathematics from an era of proof scarcity to an era of proof abundance, and that the field now needs to figure out how to verify and make sense of far more results than humans can read. His full thinking is on his website. His practical advice to students is worth reading outside of math too: only use AI on problems where you could defend the output yourself, as a way to keep your own reasoning from atrophying. It is a specific, actionable take on something a lot of people are quietly worrying about right now.