Latest Earlier stories
Report: an attacker used an open-source agent to reach sites of more than 100 companies
Reuters reviewed records of attempts on over 1,200 sites with the Cairn agent tool. The breach count is not independently confirmed.
OpenAI publishes coding evaluations for low-resource programming languages
OpenAI says it is improving ChatGPT on low-resource programming languages, with new evaluations covering software engineering, SQLite and COBOL migration.
Democrats may use midterm wins to scrutinize AI companies
Polls predict Democrats could flip at least one congressional chamber in November, which would give them committee control to press AI firms on data center tax credits and recent hacking incidents.
Meta details Prometheus, its gigawatt-scale training cluster in Ohio
It runs several hundred thousand GPUs today, aims for about a million, and supports the Muse models.
Microsoft adds per-task permissions and risk scores for agents in Microsoft 365 E7
Critical actions need human approval. One report says it is in private preview.
MIT researchers show Claude agents on a virtual island invent play-like activities
Copies of Claude Opus 4.7, nicknamed Eko, were stranded for 30 hours in a Roblox-inspired virtual island and invented games and drawings. The paper is a preprint not yet peer-reviewed, and researchers acknowledge the behavior could be role-playing.
Hugging Face publishes TorchNeuron guide for running Hugging Face models on AWS Trainium
The post covers running Hugging Face models on AWS Trainium chips through the TorchNeuron PyTorch backend, presented as a simpler path than before.
Anthropic commits 5% of its shares to a new public benefit fund
It needs stockholder approval. With earlier pledges, about 20% of shares would go to public benefit.
DeepSeek-V4.1-Flash-1017 goes live at unchanged prices
DeepSeek reports 42.6 on Humanity's Last Exam and 77.0 on DeepSWE. Its Search and Fetch API also opened to everyone.
GPT-6 Pro in ChatGPT gets faster and can split a task across parallel agents
OpenAI's own posts call the model both GPT-6 Pro and GPT-6 Astra Pro.
ChatGPT's Chrome extension can now read open tabs and fill in forms
Chat mode reads tabs. Work mode can fill forms, send messages and download files.
Meta releases Muse Spark Open weights: 816B parameters, 42B active
It reads a 1M-token context and ships under the Muse Model License on Hugging Face.
Anthropic gives six METR staff employee-level access to assess its risks
METR will publish its own analysis.
Restarting research agents from their best idea beats running them in parallel
The paper reports gains of up to 15% over parallel agents and 39% over a single agent, on four tasks.
Gemini 4 Argon is generally available at $2 in and $10 out per million tokens
Prices double on December 1. It is not on the free tier.
Meta releases Muse Wonder, its first frontier model
It costs $4 in and $20 out per million tokens and scores 52 on the Artificial Analysis index.
Anthropic says new classifiers cut what distillers can harvest by 90 to 99%
Its account names Zhipu and Moonshot among the distillation attempts.