Microsoft AI Model Development Exposes 38TB of Sensitive Data

Translated data: Microsoft recently experienced a data breach on GitHub, resulting in the leakage of 38TB of private data, including passwords, keys, and internal messages, due to the misuse of shared access signature tokens (SAS) on the Azure platform. The improper use of SAS tokens and lack of monitoring exposed the data for years, highlighting the security challenges in AI model training. This incident underscores the need for stronger security measures and collaborative efforts in the AI development process that relies heavily on large-scale data.

JetBrains Releases Programming AI Model Mellum2.1 with Twice the Throughput of Competitors Under Heavy Load

JetBrains launches a new programming AI model called Mellum2.1, focusing on enhancing agent-based programming capabilities. The model continues with a 12B mixture-of-experts architecture and open-source licensing, lowering the barrier for enterprise and individual developers to deploy privately. In terms of training, reinforcement learning has expanded from short-term completion to become the core training component, with specialized reinforcement in millions of sandbox environments for software engineering and algorithms.

SpaceXAI's Strongest Model Grok 4.7 Launches: Full Capabilities for Long Tasks, Price is Half of Flagship

SpaceXAI releases Grok 4.7, positioning it as the strongest model for programming and knowledge work, with speed twice that of competitors and a price half of the flagship version. The new version uses a larger base model, enhancing long-term tasks, self-checking, and long context management, and includes professional tasks such as documents, presentations, law, healthcare, and engineering in key testing. It aims to reduce errors and produce more reliable results in tasks taking hours.

DAMO LiON, the Liver Cancer AI Model from DAMO Academy: Detecting 1-Centimeter Hidden Tumors and Completing 15 Cases in Two Months

DAMO Academy and Shengjing Hospital released liver cancer AI model DAMO LiON, which identifies tiny lesions on contrast-enhanced CT. In a two-month prospective trial, it found 15 previously missed malignant tumors, mostly about 1 cm. Published in Nature Medicine, showing AI's value from missed to detected.....

Performance Boost! Grok 4.6 Officially Released, Standing Shoulder to Shoulder with GPT-5.6 Sol

SpaceX AI launched its flagship model Grok4.6, focusing on complex long-term agent tasks, with multiple benchmark tests matching or surpassing OpenAI's GPT-5.6 Sol. Previously, in July, it jointly launched Grok4.5 with Cursor, trained on tens of thousands of NVIDIA GB300 GPUs, and Grok4.6 further optimizes long-term complex task capabilities based on it.

OpenAI to Release New Flagship Model Astra Next Week: Largest Pre-training Since GPT-4.5, Focusing on Agent Collaboration

OpenAI reportedly plans to release Astra, codenamed mewfour, next week. It is the largest pre-trained model after GPT-4.5, marking a major capability leap. An internal version solved 10 important open math problems, with compute costs around $2,000 (~13,500 RMB), demonstrating powerful complex reasoning.....