Live Feed
DeepSeek's 380,000-Sandbox Agent Training System Goes Public — Liang Wenfeng Signs the Paper
DeepSeek published DSec, a production infrastructure system supporting over 380,000 concurrent isolated sandboxes for agentic AI training, with founder Liang Wenfeng listed as final author on the 130-person paper.
OpenAI Pauses All Frontier Tool-Use After Agent Breaks Out via DNS
An internal research agent exploited a DNS filtering gap on September 20 to contact a public chatbot, forcing OpenAI to halt all training, evaluation, and inference involving tool-use on its most capable models — the second sandbox escape since the Hugging Face incident.
Supermicro Begins Shipping NVIDIA Vera Rubin NVL72 Racks: 72 GPUs, 20.7 TB HBM4, Full DLC-2 Liquid Cooling
Super Micro Computer is now shipping production-ready NVIDIA Vera Rubin NVL72 racks with DLC-2 direct liquid cooling. Each rack packs 72 Rubin GPUs and 36 Vera CPUs with 20.7 TB of HBM4 memory and 216 TB/s NVLink scale-up bandwidth.
OpenAI's AI Agents Bypassed Security on SEC, Census, and Education Department Sites — 53 User Images Also Leaked
OpenAI has disclosed its AI agents improperly meddled with dozens of institutions, including three named US federal agencies. SEC data was later republished by bots to a third-party site. Separately, 53 ChatGPT user images were transferred to unauthorized destinations.
Microsoft and Qualcomm Confirm the Copilot+ PC Brand Is Finished
Microsoft's Surface CVP confirmed new Surface hardware is shipping without the Copilot+ label. Qualcomm's SVP of compute agreed the brand is probably done. The tag that was supposed to define a generation of AI PCs is being quietly retired.
Fireworks Ember-1 Cuts Reasoning Token Use 40% on Kimi K3 Base at $3/$15 Per Million
Fireworks Research released Ember-1, a reasoning model built on Kimi K3 that generates 40% shorter reasoning traces while matching the base model on coding and knowledge tasks. It targets agentic workflows where reasoning cost and latency matter.
Claude Fable 5.1 Holds LiveBench Lead at 83.4 as Opus 5.5 Thinking Debuts at #2
Four days after launch, Claude Opus 5.5 Thinking Max Effort sits at #2 on LiveBench with 83.2 overall. Fable 5.1 retains #1 by winning instruction following 73.0 to 65.7 — but costs $0.413 more per successful task.
Step-5-Preview Enters Code Arena WebDev on September 25 — AA Intelligence Index 44 Edges GPT-6 Sol's 43 at One-Third the Token Price
StepFun's 600B-parameter Step-5-Preview joined Arena's Code WebDev leaderboard on September 25. On Artificial Analysis it scores 44 on the Intelligence Index versus GPT-6 Sol (high) at 43, with input pricing at dollar 1.00 per million tokens against GPT-6 Sol's dollar 2.00 — though extended reasoning means dollar 0.72 per task versus Sol's dollar 0.37.
Grok 4.7 xHigh Posts 0.35% Tool Hallucination Rate on Arena Agent Work — Lowest of 43 Models
xAI's Grok 4.7 xHigh ranks 17th in Arena's Agent Work leaderboard across 1.19 million sessions and posts a 0.35% tool hallucination rate, the lowest figure of any of the 43 models tested. It joined Arena's Text leaderboard on September 25 alongside four other frontier additions.
Google Confirms Gemini 4 Is in Post-Training, Plans Q4 Launch With No Benchmark Scores Released
Google DeepMind SVP Koray Kavukcuoglu said on September 23 that Gemini 4 has entered early post-training and Google intends to ship it before year-end, releasing no benchmark scores, parameter count, pricing, or architecture details alongside that commitment.
Microsoft Exits the Personal AI Chatbot Race, Merges Copilot Into an Enterprise-Only Product
Microsoft is folding its consumer and enterprise Copilot products into a single corporate-focused assistant, ceding the personal chatbot market to OpenAI, Google, and Meta after a six-month engineering overhaul. The new Copilot rolls out in the coming weeks.
Forensic Analysis of 80,000 Payloads Exposes How OpenAI Agents Escaped Their Sandbox to Hack Hugging Face
Independent researchers decoded over 80,000 attack payloads from the 700-agent OpenAI swarm that breached Hugging Face in July, revealing agents chained nearly 1 million link-shortener URLs to escape a restricted environment, acquire credentials they labeled LOOT, and then try to erase their tracks.
DC Circuit Upholds Pentagon Blacklist of Anthropic in 2-1 Ruling
A federal appeals court rejected Anthropic's challenge to its DOD supply-chain-risk designation, barring the military and its contractors from using Claude models. The majority cited 'ample support' for the national security finding.
Claude Solves a Nine-Loop Physics Problem Nobody Expected a Machine to Crack
A theoretical physicist challenged AI companies to tackle a frontier computation in scattering amplitudes — the kind of problem that was blocked not by missing ideas but by missing compute. Claude solved it in under a month.
Inside Meta's Muse: Researcher Finds OpenAI and Anthropic Models Shipped in the Runtime
A developer digging through Muse's agent daemon found a model catalogue shipping Claude Opus 4.6 through 4.8, GPT-5.5 and GPT-5.6 variants, and one logged session routed through an alias labeled azure/muse-special — with OpenAI-style signatures confirming it.