AI
Google finds exploit_agent running fully automated attacks
Google's threat team says attackers used AI to turn patches into working exploits with no human in the loop, and more capable attacks are only months away.
This edition was produced with artificial intelligence. Text and voice are generated automatically.
Google spots first fully autonomous AI exploit pipeline
Google’s Threat Intelligence Group and Mandiant found an exposed staging server running a tool they call exploit_agent, which connects to several frontier and locally run AI models, compares popular web software code before and after each patch, writes exploits and runs them without human involvement. It is the first time Google has seen a fully autonomous attack pipeline; two months earlier its quarterly report said none had been observed. The server held stolen victim data, brute-force and webshell scripts, and configuration files with Russian-language comments. Victim data points to French media, entertainment and education sites. Google said patched flaws are easier AI targets because attackers can study old and new code, release notes and public research, and warned such frameworks will likely keep spreading.
Google says attackers used AI to turn patches into exploits with no human →
Topics: Autonomous AI agent cyberattacks in the wildAI-discovered vulnerabilities and AI bug-report slop
GPT-6 Astra helps crack 50-year-old lattice physics problem
OpenAI’s GPT-6 Astra helped produce a proof addressing a 50-year-old obstacle to putting the full Standard Model on a lattice, physicist Nathaniel Craig of UC Santa Barbara presented on OpenAI’s blog and in a preprint. Lattice simulations have calculated the proton’s mass from first principles, but the weak force treats left- and right-handed particles differently, creating ghost particles; overlap fermions can filter them out, yet for over 25 years their phase could not be set correctly. Craig had a strategy of comparing progressively finer lattices but could not prove it worked in all needed cases. GPT-6 Astra turned the idea into a working construction and led the detailed proof. Lattice theorists Maarten Golterman, Andreas Kronfeld and Yigal Shamir reviewed it and raised questions Craig says built confidence. He calls the work much more a beginning than an end.
OpenAI’s GPT-6 Astra tackles the Standard Model’s toughest math problem on a lattice →
Topics: GPT-6 Astra
Alibaba launches closed-weight Qwen 4 Max with 2.4 trillion parameters
Alibaba released Qwen 4 Max on Monday, its most powerful model and the first flagship on its new Qwen 4 architecture. It has 2.4 trillion parameters, accepts up to one million tokens of context and handles text, images and video, using a mixture-of-experts design that activates 71 billion parameters at a time, down from 95 billion for Qwen 3.8 Max. It is the highest-ranked Chinese model on Arena’s agent leaderboard, ahead of Tencent, Moonshot AI and DeepSeek, placing 14th among individual models with a preliminary 7.46% score across 14,727 sessions. On Code Arena’s web-development ranking it debuted sixth at 1,750 points. Through QwenCloud it costs $1.50 per million input tokens and $4.50 per million output tokens, 25% below its predecessor. It is closed-weight: developers can use it via API but cannot download or privately run it.
Alibaba’s Qwen 4 Max Leads China’s AI Agent Race—But You Can’t Download It →
Topics: Qwen
Nvidia releases Nemotron for Science with chemistry dataset
Nvidia introduced Nemotron for Science, combining its Nemotron 3.5 Super VL model, a chemistry dataset and post-training recipes to help researchers build AI agents for chemistry experiments, protein design and related work. The dataset, LitmusBench v0.2, contains nearly 50,000 questions for training and evaluating chemistry reasoning. Both the model and dataset are available on Hugging Face. Lila Sciences, Terray Therapeutics, Therna, Argonne, Jülich Supercomputing Centre and KTH Royal Institute of Technology are using or integrating Nemotron for scientific applications, and Riken has made the models available through its Rikyu platform. Nvidia’s life sciences portfolio also includes BioNeMo models and services for drug discovery, protein engineering and genomics analysis. Researchers and policy groups have warned that AI for biology and chemistry can create dual-use biosecurity risks.
Nvidia debuts AI model for chemistry, protein design →
Topics: Nvidia