Latest artificial intelligence news — foundation models, AI safety research, autonomous agents, and real-world AI deployments. Updated every 30 minutes from top AI publications.
A new attention architecture called Latent Sparse Attention (LSA) matches full-attention accuracy at 10M tokens while cutting attention-layer prefill time by 72.4× on an H200 GPU when distilled into Qwen-3.8-27B, accordi…
SentinelOne announced on May 21, 2026 that it integrated Anthropic's Claude platform into its security tools via Anthropic's Claude Compliance API, part of an Anthropic push that added partnerships with more than 100 sec…
Anthropic added two sentences to the Claude Code prompt on Opus 5.5 to stop the coding assistant from recreating specific existing games, according to Addy Osmani, Anthropic's engineering and evangelism leader, who confi…
An Arena.ai Agent Arena analysis of more than 26,000 agent sessions across 30 AI agents from 13 labs, covering September 12 to October 23, 2026, found top-performing AI agents process about 93 input tokens for every toke…
A developer outlined three invariants for Node.js vector collections used in media deduplication: every record in a collection must share one embedding model, one fixed dimension, and one declared chunking policy. The wr…
AiSOC, an open-source AI security operations center tool from developer beenuar, hit GitHub trending on November 2 with 446 stars in a single day and more than 6,000 total. AiSOC runs the llama3.2:3b-instruct-q4_K_M mode…
PostHog reports that its proactive analytics agents, called Scouts, changed 377 files and saved more than 3,000 broken link requests per week in the two months its 404 fixer skill has run, as the company argues AI has re…
OpenAI has added activation classifiers, called probes, to its safety stack for all assistant completions starting with GPT-5.6, the company said in a technical post. The probes are lightweight ML models that read a mode…
A developer demonstrated how pass@k and pass^k metrics answer different questions about model reliability, using four constructed Kubernetes-configuration grading outcomes to show that a model passing 50% of single attem…
Alvaro and co-founder Alex, creator of open source video-to-video tool Warpfusion, released Mago, an open source AI video studio desktop app that runs locally on a user's own GPU through a portable ComfyUI or an existing…
A developer posting as Maharishiva released tinywmcraft, a 13.9M-parameter world model of the 2D game Craftax-Classic that generates each next frame, reward and episode-end signal from pixels and actions alone, and can b…
Deep Principle released Hydrogen-1.0, a 753B-parameter open-weight scientific model for chemistry, materials science and life sciences, built by post-training Z.ai's GLM-5.3 on domain data, laboratory records and agent d…
MIT researcher Rikiya Takehi reported in a November 2, 2026 post that OpenAI's GPT-6 Astra improved Alibaba's Qwen3.5-0.8B by directly editing 0.016% of the smaller model's weights with no training run, yielding math per…
Anthropic added two sentences to Claude Opus 5.5's prompt on October 27 instructing the model not to recreate specific existing games, a rule with no opt-out flag, and users subsequently reported refusals on unrelated ta…
Prosaic Intelligence released suicide-memory-bench v01.1, a 2,048-task stateful evaluation of memory retention in production AI systems, finding that even the highest-scoring model, Gemini 4 Argon, failed to ask about us…
Meta Platforms' web crawlers have crashed and overwhelmed small-business websites, with California artist Chelsea Paulus reporting Meta-ExternalAgent generated about 550,000 hits over a few days and drove her site's CPU …
Anthropic, OpenAI, Google and Microsoft announced a new robots.txt token, __usage_ai-training, that lets webmasters block or allow all four companies' AI crawlers with two lines of code instead of listing each user agent…
Cisco reported that its SnortML machine-learning detection engine caught a novel PHP2Shell exploit attempt in customer firewall telemetry less than 24 hours after the PHP vulnerabilities were disclosed, blocking remote c…
GPT-6 Astra, a frontier LLM, can natively read and write neural network weights without gradient descent, according to research the model's developers call Weight-Space Programming. Given raw weights, Astra identified wh…
CMS sent formal notices of problems to four of the six vendors running its AI-assisted Medicare coverage review pilot, the WISeR model, according to Freedom of Information Act documents released Tuesday by the Electronic…