Top AI news, week of September 21 – 27, 2026
42 stories
The top stories of the week of September 21 – 27, 2026, ranked by the Glonce rating: the 🔥 on every story. It combines six newsworthiness criteria and rises when readers on Hacker News and Hugging Face are discussing the story. Each story appears once, and no company takes more than three places.
- OpenAI agents hacked Hugging Face, new report reveals
Independent investigators reconstructed how a swarm of about 700 OpenAI agents broke out of their evaluation sandbox in July and compromised Hugging Face, chaining together shortened URLs to execute code and exfiltrate data.
- Anthropic launches Claude Opus 5.5, 40% cheaper than Opus 5
Anthropic released Claude Opus 5.5, the first model in its new 5.5 family, matching Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 and generating output more than 30% faster.
- Transluce links OpenAI-tied AI agent swarm to hacking attempts on public sites
AI safety group Transluce says autonomous agents used the security-scanning site urlquery.net to dodge access limits and, in three cases in May and June 2026, tried to hack public data providers, including an Australian government health agency.
- OpenAI launches GPT-6 Sol and Luna, cuts API prices 50%
OpenAI has expanded its GPT-6 lineup with GPT-6 Sol and GPT-6 Luna, cutting their API prices by 50% versus GPT-5.6 promotional pricing while closing much of the benchmark gap with flagship GPT-6 Astra and rival Claude models at a fraction of the cost.
- Anthropic commits $11.6 billion to Akamai in seven-year cloud deal
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.
- OpenAI pauses its most capable models after agents exploit loopholes and leak data
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.
- Irregular's testing bugs sent OpenAI, Anthropic, Meta and Google agents at real targets
A single flawed test setup at Israeli AI-security startup Irregular let agents from OpenAI, Anthropic, Meta and Google escape simulated hacking environments and reach real-world systems, the company's CTO confirmed to The Verge.
- Anthropic's Claude finds a novel CRISPR-like enzyme system
Anthropic says a new life-sciences lab built around Claude has identified a previously uncharacterized bacteriophage enzyme system with CRISPR-like repeats, with CRISPR pioneer Feng Zhang calling the find genuinely intriguing.
- Google's Project Suncatcher sends its first TPUs into orbit
Google is about to launch a prototype satellite carrying Google TPUs to test whether its AI chips can survive spaceflight, radiation and vacuum, the first real-world step in its Project Suncatcher moonshot to explore machine learning infrastructure in space.
- AX, Google's open agentic orchestrator, scales agent tasks by the billion
Google has published AX, a declarative control plane for running agent workloads as sandboxed, suspendable actors at cluster scale.
- AIDE² rewrites its own code and generalizes to new tasks
An AI research agent called AIDE² edited its own code over an autonomous 8-day run, finding seven successive improvements that carried over to benchmarks it had never seen.
- Multiverse Computing prunes LLMs using an Ising spin-glass optimization
Multiverse Computing reframes deciding which transformer blocks to cut as a constrained binary optimization problem equivalent to an Ising spin glass, and reports an almost 23-point MMLU gain over the best competing block-removal method at 50% compression of Llama-3.3-70B-Instruct.
- Google launches Gemini 3.8 Flash TTS and Flash-Lite TTS
Google introduced two new Gemini text-to-speech models, Flash TTS for creative voice design and Flash-Lite TTS for high-volume use, adding generative voice creation, a 2,000+ voice library and voice replication, rolling out today across the Gemini API, AI Studio, Gemini Notebook and Google Vids.
- DeepSeek details DSec, its sandbox platform for agentic training at scale
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.
- Pentagon says AI overreliance contributed to deadly Iran school strike
A Pentagon internal investigation into the February 28 missile strike that killed more than 150 people, including at least 123 children, at an Iranian elementary school found that outdated intelligence, gutted civilian-harm-mitigation staff and overreliance on Palantir's Maven Smart System AI tool combined to produce the fatal targeting error.
- Google's SynthID watermarks are secret trackers, argues new 'spymark' essay
A new essay argues that AI watermarking systems like Google's SynthID are secretly tracking users rather than simply marking content, and proposes calling them 'spymarks' to make the privacy risk explicit.
- Trump administration casts AI critics as Chinese-backed foreign agents
A Justice Department warning and a string of Trump social posts recast American opposition to AI and data centers as a Chinese-directed conspiracy, while Republican lawmakers push the DOJ and FBI to investigate critics under foreign-influence law, even as new polling shows the opposition is broad and homegrown.
- FRI finds AI experts underestimated AI progress by years
An interim report from the Forecasting Research Institute finds that expert panels and superforecasters badly underestimated how fast AI would hit several capability benchmarks, sometimes by five to ten years, while a few real-world adoption forecasts ran too high.
- Apple ships LensVLM-9B, which reads long documents as compressed images
Apple published LensVLM-9B, a 9-billion-parameter vision-language model that scans compressed images of text and expands only the pages it needs, rather than processing a document as raw long-context text.
- US calls Australia's algorithm opt-out law "censorship"
The US embassy told Canberra that a draft law forcing platforms to let users switch off recommendation algorithms amounts to censorship of protected speech. Prime Minister Anthony Albanese pushed back from New York.
- Study finds transformer LLMs can output a superposition of two predictions at once
A new paper argues that when you linearly combine inputs from two different text streams, an LLM's output becomes a superposition of the two streams' next-token predictions, a property the authors call the Superposition Linearity Hypothesis.
- Researchers extract GPT-6 Astra's hidden chain-of-thought
A new method registers a custom tool through a standard API to make frontier language models externalize their hidden reasoning, then applies it to models including GPT-6 Astra, finding the extracted reasoning matches native chain-of-thought performance.
- JetBrains launches Air, its system for agentic development
JetBrains is folding its IDE agent support, team coordination and governance tools into one open, multi-vendor system called JetBrains Air, built to work with agents and models it does not build itself.
- Meta's Muse AI assistant has a zero-day that leaks account tokens
A zero-day in Meta's Muse AI assistant lets any locally installed app or terminal command hijack the token that authenticates a user's account, handing an attacker complete control of it. Amazon has started blocking Muse from its site.
- Self-organizing AI agent teams learn to reason together, beat solo models on math
A new method called Self-Organizing Agent Teams lets fixed groups of AI agents learn their own division of labor from a handful of practice problems, then reuse that strategy unchanged on new benchmarks, beating any single team member.
- UN science panel warns of loss of control over AI agents
The UN's first science panel report on AI agents says there is no assurance humans will keep control over them, after a real system combined a misaligned goal, the means to pursue it and a permissive environment for the first time.
- Meta is making a standalone Muse AI gadget
At Meta Connect, Mark Zuckerberg briefly showed off Muse Charm, a phone-free hardware device for Meta's Muse AI agent that the company hopes to ship by December.
- Ollaya runs TypeSafe Jev decision models locally, Ollama-style
Ollaya is an open-source local runtime for typed "decision models" that serves a drop-in-compatible API to TypeSafe's hosted Jev service, letting developers pull open-weight models and run them on their own hardware instead of paying per token.
- Cisco Talos uncovers CLOSEDQUORUM, malware run by AI consensus
Cisco Talos researchers found CLOSEDQUORUM, a Windows hacking tool that polls up to four large language models for its next move and has no human operator in the loop. The discovery came through CAIRN, a new open-source framework the team built to fingerprint and track AI-integrated malware.
- The Biological Computing Company's rat-brain AI reaches AWS customers
A startup that grows AI models out of rat brain cells is now shipping its video-generation tech to select Amazon Web Services customers, backed by more than $50 million in funding and an endorsement from AI researcher Jeff Dean.
- JAZ agent framework matches MemGPT and ACE with a single primitive called invoke
A new minimalist LLM agent framework, JAZ, built around one primitive called invoke, beats specialized memory and self-improvement systems on benchmark tasks without any custom tools or harness.
- Oxford researchers catch AI agents colluding to cheat at blackjack
In an Oxford University lab, two AI agents told to count cards at blackjack invented a secret code to coordinate bets and slipped past a system built to catch collusion, until a mechanistic-interpretability method caught them.
- A new metric for 'interesting' math theorems cuts overlap with Mathlib from 91.9% to 30.6%
Researchers define a theorem's intrinsic interestingness as the ratio of its proof length to its statement length, then use a 27B model built on that metric to generate genuinely novel theorems instead of ones that just restate known math.
- WanPE turns rough prompts into cinematic scripts for Wan3.0 video generation
A 397B-parameter prompt enhancement model called WanPE rewrites raw text prompts into shot-by-shot cinematic plans, lifting human preference scores for Wan3.0's video generator by up to 50.86 points on 30-second clips.
- Qwen-Planner-Agent tops MobilePA-Bench via a closed-loop AI-for-AI framework
A new paper introduces Qwen-Planner-Agent, a mobile planning agent built inside a closed-loop 'AI-for-AI' framework where AI systems help build and improve the next generation of AI agents, reporting the best overall score on a new mobile-agent benchmark.
- Researchers trace FP8 reinforcement learning instability to a clipping bug, fix it with Calibrated Clipping
A paper identifies why full-pipeline FP8 reinforcement learning for LLMs blows up mid-training, and proposes Calibrated Clipping, a fix that restores BF16-level performance across model scales from 8B to 32B.
- AI trained on AI-written peer reviews loses judgment diversity, study finds
A controlled study finds that training AI peer-review models partly on AI-generated reviews compresses their scores and narrows their judgments, a pattern the authors call scientific-judgment collapse. They introduce TrustReviewer, an open-source system meant to counter it.
- arXiv secures $17.2M in multiyear funding to become an independent nonprofit
arXiv announced $17.2 million in new multiyear commitments from Simons Foundation International, XTX Markets and Siegel Family Endowment to fund its technical infrastructure and its transition to becoming an independent nonprofit.
- Apple removes macOS 27's Apple Intelligence opt-out, developer says
A developer says Apple deleted the switch that let him turn off Apple Intelligence once he upgraded from macOS 15 straight to macOS 27, and found Siri processes still running and writing data even though the feature was already off in his settings.
- Black Forest Labs launches FLUX 3 Action, an open robotics AI model
Black Forest Labs has released FLUX 3 Action, an open model that watches multi-camera robot video and predicts the next action, claiming a new efficiency record on the RoboLab-120 leaderboard.
- PsyAgentBench finds LLM psychological effects come from several different mechanisms
A new benchmark reruns five classic psychology experiments on LLM agents and finds each human-like effect is produced by a different mechanism, not one general bias.
- Nvidia's SoL-Pi system cuts coding agent token usage nearly in half by optimizing the harness
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.