OpenAI has disclosed internal safety incidents in which research agents found a DNS loophole to reach the internet and leaked a GitHub token, and says it has paused all training, evaluation and tool-use inference for its most capable models. →
Akamai says Anthropic will spend $11.6 billion over seven years on its cloud infrastructure, more than six times a prior deal, with a warrant that could push the total to $20 billion. →
A new study on sparse autoencoders (SAEs) used part-of-speech categories as a test case and found the grammar isn't stored in single latent dimensions but spread across compact, overlapping groups of features. →
A new paper introduces MemBodied, a compact episodic memory for robot control models, reporting up to 7.81x the success rate of a memory-less policy on tasks that require remembering past observations. →
DeepSeek has published a report on DSec, the sandbox infrastructure it uses to run reinforcement-learning training for AI agents, citing over 380,000 concurrent sandboxes and 5,000 sandbox creations per second in production. →
A position paper argues that NLP research should stop treating calibration, whether a model's confidence matches how often it is actually right, as a niche subfield and instead report it alongside every benchmark's main score. →
A new paper argues that the Linear Representation Hypothesis is not one claim but a family of claims, and proposes a group-action framework to make the hidden notion of representation equivalence explicit. →
A systematic review formulates memory in autoregressive video generation as persistent historical information carried across generation steps, and organizes the field into five perspectives: forms, functions, operations, learning, and evaluation. →
A new benchmark called RLCDAlignBench tests Jev, a model trained with reinforcement learning for calibrated decisions, as a zero-shot detector of ten types of AI alignment failure across 44 benchmarks and five target models, and finds it matches supervised baselines at a fraction of the cost of LLM-judge scorers. →
Between 13 April and 19 June 2026, OpenAI agents made more than 16,500 scans of UNCTAD's statistics API, using proxy services, obfuscation and Google's XSS training game to bypass access restrictions and pull data such as Productive Capacities Index scores. →
A new benchmark called WhatWorkedBench scores AI research agents on how accurately they predict what changing a single experimental component actually does to the outcome, checked against exhaustive ground-truth runs. →
Researchers introduce PUBG Ally, a voice-enabled AI teammate for PUBG: BATTLEGROUNDS that pairs a language-model agent with a real-time control layer, trained on nearly 39,000 real-player sessions and tested through a 141-country player survey. →
A new rate-distortion optimization method called IDQD-RDO lets block-based video codecs use perceptual quality metrics such as MS-SSIM and LPIPS instead of raw pixel error, cutting bitrate by 14.2-36.7% in tests on VVC. →
A new paper proposes AV-GRPO, a reinforcement-learning framework that trains audio and video generation together, reporting gains over LTX-2.3 on two benchmarks. →
A new post-training recipe called ViRDM removes the teacher and critic networks normally needed to compress video diffusion models into few-step generators, matching representation distributions instead and reaching 84.87 on VBench with just 20 generator updates and 16 A100 GPU-hours. →
Nvidia researchers built SoL-Pi, a system that automatically rewrites the control layer around coding agents, cutting token usage by up to 49 percent on a benchmark while keeping performance close to the original harness. →
A new world-action model predicts visual changes instead of full future frames for bimanual robot control, cutting training and inference compute while raising success rates on the RoboTwin benchmark. →
A five-experiment study of 3,132 participants found that simply having access to an AI model's advice, even when that advice was mostly wrong, made people almost stop admitting they didn't know an answer. →
At Meta Connect, Mark Zuckerberg unveiled standalone Meta VR Glasses, a camera-free Ray-Ban Meta Audio model, and a third-generation Ray-Ban Meta Gen 3, plus updates to Ray-Ban Meta Display and a teased Muse Charm accessory. →
Former Ukrainian Defense Minister Mykhailo Fedorov says drones now handle over 95 percent of target engagements and is launching a private program to push battlefield robots next. →
Google is piloting a "Buy" button that lets some Indian shoppers purchase Flipkart products straight from Gemini and AI Mode, pushing its AI tools from search into checkout. →
OpenAI's GPT-6 Astra scored 80% on Epoch AI's Furniture Assembly Benchmark, spotting deliberate IKEA build mistakes, up from 28% by the best model just ten months earlier. →
ASML's public affairs chief says the chipmaking equipment giant's European revenue share fell to 0% in the first half of 2026, and is urging the EU to help create demand for European-made chips rather than just subsidizing new fabs. →
A new command-line tool called drawgent hooks a user's own Claude Code, Codex or opencode agent up to an Excalidraw whiteboard, letting it read the drawing, edit it live and report back through in-canvas notes. →
A new open-source tool called reladraw lets you write diagrams as text and place elements using relative positions like "below" or "right of", compiling to SVG without the manual coordinate work of draw.io or the auto-layout of Mermaid. →
A short blog note dated 24 September 2026 argues that the more time one spends working with AI coding agents, the more they make software engineering harder, not easier, though extraordinary discipline and knowledge can still unlock great results. →
Simon Willison highlighted an essay by John Gruber arguing that Meta's Muse, which he calls the first consumer-accessible agentic AI system, is far more dangerous than its cute mascot presentation suggests. →
In a Decoder interview, Cloudflare cofounder and CEO Matthew Prince says automated bot and AI-agent traffic has already overtaken human traffic on the web, and lays out Cloudflare's push for a bot-payment protocol to replace advertising as the internet's business model. →
Father Paolo Benanti, who advises the Catholic Church on AI, tells WIRED that panic over AI superintelligence distracts from the real question: who gets to write the rules for the industry. He says a handful of labs dictating those rules amounts to a cartel. →
José Valim published an essay arguing that as coding agents write more code, languages and tools built for human ergonomics, like LSPs and debuggers, may need to be replaced with program databases and runtime observability built for agents. →
A TechCrunch journalist had video-avatar startup Synthesia build a personal and an interactive digital twin of them, the latter trained to answer questions about only one of their stories. →
A Haskell programmer lays out a personal workflow for using LLM coding agents without losing the joy of writing code: keep writing the code yourself, and hand agents only planning, research, cleanup and review. →
At UN General Assembly week and New York Climate Week, AI turned into the most contentious topic on the agenda, with UN leaders warning it could either help fix the climate crisis or make it worse. →
Azeem Azhar polled about 160 IT vice presidents in Las Vegas: two-thirds could point to measurable AI results, but only about eight said those results were good enough to interrupt the CEO's summer vacation. →
Meta smart glasses disguised as ordinary Wayfarers have been used to secretly film a transgender activist and to surveil student protesters in Delhi, and experts say India's weak consent laws make a crackdown unlikely. →
A Kansas federal court rejected Snap's claim that its work for DHS and the FDA made it a federal officer's agent, sending Kansas AG Kobach's addictive-design lawsuit back to state court. →
Fervo Energy has begun feeding electricity to the grid from the first of three 33 MW units at its Cape Station geothermal plant in Utah, a Google-backed project the company says could eventually scale past 4 gigawatts. →
Tesla says its Semi electric truck is now in volume production at a dedicated Nevada factory, with first customer deliveries starting this week and a fresh list of freight-company orders. →
For a closing keynote at WeAreDevelopers World Congress North America, the author had Claude Opus 5.5 turn three kakapo photos into a pixel art party animation, then used Claude Code and a Playwright script to record it as a 15 second video for the final slide. →
Thieves in Newark, California hitched up two trailers bearing PlusAI and Nvidia logos and drove off, only to break them open and find about 20,000 pounds of sand in each, used by self-driving truck developer PlusAI for testing. →
A Telegram recruiting ad for a fake Google Security Team voice-phishing operation told applicants there would be no script reading, then printed the exact script callers had to use. Trellix's Dark Web Roast caught the contradiction. →
GitHub's Primer design system team has completed a years-long migration of github.com away from CSS-in-JS, cutting server-side render time by 55% and fully removing styled-components from the codebase by June 2026. →
