From a new frontier coding model out of xAI to a sharp privacy investigation targeting OpenAI’s ad platform, today’s AI headlines span the frontier of model releases, agent infrastructure, and data-privacy scrutiny. Here are the five stories driving the conversation this morning.

xAI Ships Grok 4.7, Its Most Capable Coding and Knowledge-Work Model

SpaceXAI announced Grok 4.7 on September 21, describing it as its most powerful model yet for coding and knowledge work. The model is “twice as fast, at half the price of comparable models,” and is served at the same price and speed as Grok 4.6: $2 per million input tokens and $6 per million output tokens, with a fast variant at twice the output speed for twice the price.

Under the hood, Grok 4.7 uses a new, larger base model than 4.6 and was trained with a longer reinforcement-learning run weighted toward problems that take many hours to complete. It is better at verifying its own work, managing longer context, and natively understands the Grok Bot harness. On benchmark comparisons, Grok 4.7 hit 46.3% on CursorBench 4.0 (versus 40.4% for Grok 4.6), 71.0% on DeepSWE v1.1, and 64.0% on the electrical-engineering EEBench.

The release also debuts an entirely new safeguard stack, leading on refusal rates and jailbreak resistance — including the strongest score xAI has seen on LatchBio’s biosafety benchmark at 62.4%, and allowing just 3.3% of risky dual-use prompts through on HackerBench v0.3. Grok 4.7 is available today in Cursor and Grok Build, through the Grok API, and via model routers and cloud platforms.

Google’s AX: An Open Agentic Orchestrator for Scaling Agent Workloads

AX (agentexecutor.io) is an open-source agentic orchestrator built by Google engineers, and it climbed to the top of Hacker News this week with more than 640 points. Its pitch: “Declare an agentic task. AX runs it at scale,” sandboxing each task, wiring up its workspace, and fencing its network so you can run many agents per cluster.

The project centers on four declarative primitives — Task (isolated execution with CPU/memory limits), Workspace (easy Git/MCP/skills setup), and related constructs — so that untrusted agent code runs in a sandbox that is cheap to create, suspend, and throw away. Workflows are defined in plain YAML files and managed with the ax CLI.

Commenters noted the natural fit with Google’s broader agent tooling, such as Antigravity and Jules, and welcomed an open option for the growing stack of agent sandbox and orchestration startups. Others were quick to caution that while it was developed by Google employees, the project does not necessarily carry full official Google backing.

Investigation: OpenAI’s Ad Collector Ties Your Web Browsing to Your ChatGPT Account

A detailed investigation published this week alleges that OpenAI’s ad platform connects what you do on ordinary websites to your ChatGPT account via an identifier called __obi. The mechanism, documented at bzr.openai.com (OpenAI’s internal “bazaar” ads system), begins when ChatGPT generates a JWT that binds a stable identifier to your account, then sets the __obi cookie scoped to .openai.com.

Advertisers that run OpenAI ad pixels load a small SDK that transmits the cookie — along with page data such as products searched, articles read, and purchase behaviors — back to OpenAI’s servers. The author says they reproduced the full mechanism on their own phone, verifying it with two independent capture methods and cross-checking months of traffic spanning 936 distinct advertiser pixels across 1,029 hostnames, including Chewy, Wayfair, ThriftBooks, Eventbrite, HelloFresh, Coursera, and SeatGeek.

Even when logged out, an anonymous identifier per device was observed persisting at least 27 days. The report also notes the SDK harvests identity from form fields and tag-manager buses, hashing email and phone, while sending some geo data in the clear. OpenAI’s cookie policy lists __obi as an analytics cookie with a one-year lifespan. The story drew sharp community reaction over the precedent of running “standard adtech” inside an AI chat product.

Kev: Small, Self-Trainable Decision Models Built on Qwen3.5

Kev is a family of small “Jev-like” decision models you can train and run yourself. Released by Jared Palmer, the project provides 0.8B, 4B, and 9B models built on Qwen3.5 and based on the architecture described in “Jev’s Architecture Unmasked,” with full training code, frozen evaluation suites, and Apache-2.0 licensing.

Kev answers yes/no, multiple-choice, and rating questions in a single request, with questions sharing the input text but kept isolated from one another. It runs on CUDA, ROCm, and Apple Silicon — the 4B and 9B models fit a 32GB Mac using bf16 — and its API matches TypeSafe’s System One, so their Python SDK can point at your local server. A Hugging Face Spaces demo lets you try Kev-4B and Kev-0.8B in the browser with no install.

The project’s popularity reflects a wider community appetite for small, open, locally-run models, though commenters debated whether fine-tunes on RLHF-trained Qwen base models can truly be called “Jev-like,” given Jev’s own reliance on RLCD training.

macOS 27: Users Seek Workarounds to Avoid Downloading AI Models

Early macOS 27 testers are hitting a storage surprise: the OS downloads large foundational AI models locally to power the rebuilt Siri and other on-device features. A workaround posted to the macOS Beta subreddit — showing users how to prevent those downloads and reclaim disk space — drew more than 220 points on Hacker News and lightened a debate about user choice.

Commenters split between those eager for the new on-device Siri (noting it is handy for search and local actions) and those frustrated at the lack of an explicit opt-out, with several saying they will hold off upgrading until Apple offers a real choice. Worth noting: the foundational models can be used by more than Siri — apps, shortcuts, and other tools can call them for local inference — so the trade-off is between local capability and hundreds of megabytes of storage.

That’s today’s slice of the AI world — from frontier model benchmarks to a privacy deep-dive on OpenAI’s ad stack, small open models, and the storage realities of on-device AI.

Monday, September 21, 2026, brought a characteristically busy day in artificial intelligence: Google surfaced an open-source orchestrator for running AI agents at industrial scale, OpenAI faced fresh scrutiny over a tracking cookie tied to its new advertising business, SoftBank moved to finance an additional $10 billion OpenAI stake with junk-rated debt, and researchers demonstrated that Anthropic’s Claude could be weaponized to break into OpenAI’s own internal systems. State governors, meanwhile, renewed their push for federal guardrails. Here are the five stories that defined the day.

1. Google Unveils AX, an Open-Source Agentic Orchestrator

Google captured the top of Hacker News on Monday with AX, an open-source (Apache 2.0) project it describes as “Google’s open agentic orchestrator.” AX is a declarative control plane for running AI agent workloads at scale on Kubernetes, built on a runtime Google calls Agent Substrate, which traces back to Google DeepMind’s agentic-runtime research.

AX reframes agent workloads as their own class of compute — short-lived, bursty, frequently idle while waiting on a model or a human — and introduces four primitives: Task (isolated, sandboxed execution with CPU and memory limits), Workspace (the repos, MCP servers, and skills an agent needs, or a “generative workspace” where the agent sets up its own environment from a plain-English goal), Gateway (an explicit network-policy allowlist plus credential injection), and Model (centralized model, parameter, and secret configuration). Headline claims include billions of concurrent agent tasks per cluster and sub-second resume from suspend with no cold start.

The project sits directly under Google’s official GitHub organization — a signal several commenters took as evidence of institutional backing, though the thread also raised the skeptical precedent of Gemini CLI, which was effectively superseded by Antigravity. Comparably positioned projects include Google Cloud’s Scion and OpenAI’s Agents API, which entered public beta earlier this month. AX is at github.com/google/ax.

2. OpenAI’s Ad-Tracking Cookie Triggers a Privacy Backlash

After OpenAI announced that its ChatGPT advertising business had reached a $1 billion annualized revenue run rate and was expanding globally, a privacy researcher’s findings fueled the day’s most-discussed OpenAI story. Investigators documented an ad-measurement pixel at bzr.openai.com that mints a JWT-bound cookie scoped to .openai.com, configured with SameSite=None and a one-year TTL.

Because the cookie is set to travel across sites, any advertiser that embeds OpenAI’s pixel can ping OpenAI with the visitor’s ChatGPT-linked identifier. The researcher reported seeing it appear on consumer sites including Chewy, Wayfair, ThriftBooks, Eventbrite, HelloFresh, Coursera, and SeatGeek. The disclosure drew a large Hacker News thread — topping 700 points and more than 370 comments by late Monday — with critics asking how cross-site tracking squares with OpenAI’s privacy promises as ads roll out to free users.

3. SoftBank to Raise Over $11 Billion in Junk Bonds to Fund OpenAI

In what would rank among the largest high-yield bond offerings ever, SoftBank Group launched a bond sale exceeding $11 billion to deepen its bet on OpenAI, according to Bloomberg and a reported term sheet. The package comprises $10 billion of U.S. dollar notes across three maturities and €1 billion (about $1.1 billion) of euro notes across two tenors.

Proceeds are earmarked largely for SoftBank’s ~$10 billion payment on the third tranche of its follow-on investment in OpenAI — a transaction expected to close on or around October 1 — plus refinancing of an existing bridge loan. The deal is underwritten by Citigroup, Goldman Sachs, JPMorgan, and Morgan Stanley. SoftBank has now committed nearly $65 billion to OpenAI, tying an ever-larger share of its financial future to the ChatGPT maker and highlighting the growing role of debt in financing the AI boom.

4. Researchers Use Claude Opus 5 to Hack Into OpenAI’s Internal Systems

The three-person security firm Hacktron AI — researchers Harsh Jaiswal, Mohan Pedhapati, and Rahul Maini — disclosed that they used Anthropic’s Claude to compromise multiple OpenAI employee ChatGPT and Codex accounts and reach OpenAI’s internal code repository in under 72 hours.

Working from OpenAI’s public discussion forum, the team chained two flaws: a heap buffer overflow in the libheif image library, reachable through HEIC/HEIF uploads and exploitable for remote code execution, and a weakness in OpenAI’s single sign-on under which forum-issued session tokens remained valid for ChatGPT and Codex — some belonging to employees. Notably, the exploit succeeded only after Anthropic released Claude Opus 5 (July 24); the researchers said Opus 4.8 failed across multiple sessions. To prove access without examining sensitive code, they had a compromised Codex account open a pull request in OpenAI’s internal monorepo. OpenAI said both issues are now resolved and thanked the team, which reported the findings in July.

5. Governors Moore and Pritzker Push for Federal AI Guardrails

Two Democratic governors took their case for national AI oversight to the Sunday shows. Maryland Gov. Wes Moore, who chairs the National Governors Association, told CNN’s State of the Union that state efforts “cannot happen in the absence of federal leadership,” adding, “Having 50 different guidelines is not the most productive thing for us to be doing and thinking about.”

Illinois Gov. JB Pritzker echoed the theme on ABC’s This Week: “My great concern, though, is you can’t do this state by state… this needs to be a national effort as well as an international endeavor, especially together with our allies, because AI is not going to go away.” Pritzker signed Illinois’ Artificial Intelligence Safety Measures Act, which requires audits of frontier models, while the pair’s comments position them against the administration’s hands-off approach and follow warnings from Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, and Elon Musk that the technology needs to slow down.

That is the AI landscape for Monday, September 21, 2026 — from Google’s infrastructure play and OpenAI’s privacy questions, to a record AI-financing deal, a striking security disclosure, and a deepening political debate over who sets the rules.

Artificial intelligence continued to dominate the news this week with revelations about industry giants, a near-miss inside the U.S. military, and fresh demonstrations of what AI can do — from cracking a century-old cipher to designing its own silicon. Here are the top five AI stories of the day.

Microsoft exec called AI scraping “the largest theft of labor in human history”

Newly unredacted filings in the copyright lawsuit The New York Times brought against OpenAI and Microsoft in 2023 have surfaced an explosive admission: internally, the companies themselves described their AI training practices as theft. According to TechCrunch, a top Microsoft executive privately characterized the practices as “an astonishing theft of unprecedented proportions” and “the largest theft of labor in human history.”

The documents detail how the two firms allegedly bypassed paywalls undetected, mass-scraped news content into training datasets, and deliberately stripped copyright notices from training data. OpenAI’s own leadership reportedly conceded that its models posed an “existential threat” to the publishers whose work trained them. OpenAI’s mid-training datasets alone contained more than 91,692 copies of works published by the NYT, the Daily News, and the Center for Investigative Reporting, while one Common Crawl-derived dataset included more than two million documents from nytimes.com.

Microsoft CEO Satya Nadella testified in a deposition this year that paywalled content should be licensed by anyone who wishes to use it for training or grounding, and said he would have invoked Microsoft’s right to require OpenAI to retrain its models had he known the firm scraped paywalled content. Microsoft’s own data showed its Copilot “answer engine” drove New York Times click-through rates down as much as 93% versus traditional Bing search — a “doom loop” an internal Microsoft presentation said would “hurt the performance of our models and the entire web at the same time.” Both companies declined to comment.

U.S. military had close call after AI hallucinated an intelligence report

An intelligence report circulated across the U.S. military earlier this year nearly triggered an armed intercept of a Chinese ship — until officials discovered the report had been generated with the help of AI and contained false conclusions. CNN, citing four sources, reported that the report claimed a Chinese ship in the Middle East was transporting components of a nuclear weapons program, prompting plans to board the vessel with armed personnel and military aircraft in the air.

Before the operation, officials dug deeper and found a special operations analyst had used a chatbot that inaccurately identified the cargo. The report was described by one source as “entirely false” but one that “almost started a war.” An analyst queried a chatbot about intelligence originating with U.S. Special Operations Command Pacific, based in Hawaii; the bot fused open-source intelligence with signals intelligence before reaching its conclusion, and the analyst then used AI again to package the findings into a standard intelligence report. Officials say the episode underscores the profound risks — and the danger of catastrophic miscalculation — as the military pushes AI into targeting and analysis during wartime.

GPT-6 Astra solves a previously unsolved WWI German radio cipher

The GPT-6 Astra model has cracked a German World War I radio message that had eluded codebreakers for over a century, according to a writeup on Prinz. The message, transmitted November 27, 1918, was encoded using the ADFGVX method — a cipher built on a letter table assembled around a keyword. Hundreds of such messages had been decoded by experts like George Lasry, but more than a dozen had remained unsolved.

Astra determined that the encryption keyword was “TRUPPENVERSCHIEBUNG” (German for “troop deployment”) — notable because that key is believed to have been in use starting December 9, 1918, after the message was sent. The decoded text reads: “An English cruiser arrived at Sevastopol on the 24th. An allied squadron follows on the 26th.” The model verified its work against naval logs: the British cruiser HMS Canterbury did arrive at Sevastopol on November 24, 1918, and an allied squadron followed on November 26. It’s a small but striking demonstration of what modern models can accomplish in cryptanalysis.

OpenAI used its own LLMs to design its Jalapeño chip — and it was fast

OpenAI’s debut AI accelerator chip, Jalapeño, was designed with heavy help from OpenAI’s own large language models, IEEE Spectrum reported. The chip delivers up to 13.4 petaflops of 4-bit compute, accesses 232 gigabytes of HBM4 memory at up to 15.4 terabytes per second, and can reduce end-to-end inference latency by up to 3.6 times versus Nvidia’s GB300 while drawing less power.

The design went from first concept to first silicon in under 20 months, with just nine months separating the first register-transfer level (RTL) code from tape-out. “The models are giving superpowers to our engineers,” said OpenAI vice president of hardware Richard Ho. The team averaged fewer than 100 people over the project, with OpenAI handling end-to-end system design and partner Broadcom managing physical design. A front-end workflow built around Google’s open-source Accelerated Hardware Synthesis (XLS) toolchain let engineers write in DSLX and C++ before conversion to Verilog. Experts called OpenAI’s timeline “likely best in class today,” while cautioning that Broadcom’s help was essential to the pace.

Alibaba open-sources RADAR, a medical AI that detects cancer and nearly 150 conditions

Alibaba has open-sourced RADAR, a generalist vision-language model for radiology that its developers say can detect cancer and nearly 150 conditions from abdominal CT scans. According to the model card on Hugging Face and reporting from the South China Morning Post, RADAR was trained on over 400,000 contrast-enhanced abdominal CT examinations with 15 million anatomy-aware image–text pairs, learning directly from clinical reports without manual annotation.

The model is described as an “expert-level generalist AI for abdominal CT diagnosis,” capable of handling both routine and complex clinical tasks, and is released under a CC BY-NC-SA 4.0 license with code available on GitHub. The work is detailed in a paper in Science and follows a supporting study in Nature Medicine on large-scale AI-guided liver malignancy diagnosis. While experts caution about how diagnostic metrics are evaluated in practice, the release marks another significant open-source entry from China in the medical AI space.

From courtroom revelations over training data to a military near-miss and breakthroughs in cryptanalysis and chip design, today’s AI news shows a technology advancing rapidly — and raising urgent questions about how it should be governed.