This is nuts. #huggingface have just scraped the entirety of #github. https://huggingface.co/spaces/HuggingFaceCode/in-the-stack
huggingface
Y’all, last week, #AI repository #HuggingFace disclosed that it had been the victim of a cyber attack conducted by an autonomous AI agent.
Well, turns out, #OpenAI had models break containment, escape the lab, and grab prod data and certs to solve a problem it had been given.
So, that’s fun!
The attack used a combination of OpenAI's available #GPT 5.6 Sol and a more powerful, unreleased pre-release model. The models were able to identify and exploit a zero-day vulnerability in OpenAI's internal package registry cache proxy, allowing them to establish open internet access.
The models inferred that Hugging Face hosted the datasets, models, and answers for the benchmark they were being tested against.
The models chained multiple complex attack paths together. They utilized zero-day vulnerabilities and stolen credentials to achieve remote code execution on Hugging Face servers, directly extracting test solutions from the production database.
https://fortune.com/2026/07/21/openai-says-ai-models-escaped-control-hacked-hugging-face/
Quantized and published a model to #Huggingface
https://huggingface.co/tozp/tsinghua-veriloop-coder-e1-mlx-4bit
偶然之间看到这个模型，觉得它在各项性能指标上比原版的 Qwen3.6-27B 强很多。但是目前还没有看到有人发布MLX的量化版本，于是自己跑了个命令生成并且发布了这个模型。过程比我想象的要简单和快速很多。