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.

The week delivered a concentrated reminder that AI’s biggest open questions are no longer just technical — they are legal, security, and national-security questions. Freshly unsealed court filings expose what Microsoft executives privately called “the largest theft of labor in human history,” researchers revealed chain of vulnerabilities that briefly reached inside OpenAI’s internal repositories, and a U.S. military intelligence report generated with the help of a chatbot nearly prompted a boarding operation on a Chinese vessel. On the build side, a new compression breakthrough brought 27B-class capability into a 5.9GB footprint, while a deep-dive reverse engineering of Z.ai’s ZCode raised serious questions about what AI coding assistants silently ship to the cloud. Here are the top AI stories for September 19, 2026.

Unsealed Filings: Microsoft Exec Called AI Scraping “The Largest Theft of Labor in Human History”

Newly unredacted information in The New York Times‘ copyright lawsuit against OpenAI and Microsoft reveals that a top Microsoft executive privately described the companies’ AI training practices as “theft” — and that OpenAI’s own leadership said its models posed an “existential threat” to publishers and journalists. The unsealed material, reported by TechCrunch, details how the firms allegedly obtained content by bypassing paywalls undetected, building training datasets through mass scraping, and deliberately stripping copyright notices from training data.

The filings undercut OpenAI’s central fair-use defense. Microsoft’s own data shows its Copilot “answer engine” cut click-through rates for The New York Times’ domain by as much as 93% compared with traditional Bing search. An internal Microsoft presentation by director of Applied Science Brent Hecht called it a “doom loop.” The documents reveal OpenAI’s mid-training datasets alone contain more than 91,692 copies of works published by the NYT, Daily News, and Center for Investigative Reporting, while a Common Crawl-derived dataset held more than 2 million documents from nytimes.com. In a January 2023 memo, Hecht described it as “an astonishing theft of unprecedented proportions” and “the largest theft of labor in human history.”

OpenAI CEO-adjacent figures and Microsoft’s Satya Nadella have weighed in through depositions and internal messages. OpenAI’s head of ChatGPT, Nick Turley, wrote that publishers face an “existential threat” from largely substitutive products, and OpenAI President Greg Brockman described the models as “excellent at news.” The case continues with no clear resolution on whether AI training constitutes fair use, though judges have so far leaned toward AI companies’ arguments.

Researchers Chain a Heap Overflow and SSO Flaw to Reach OpenAI’s Internal Repositories

Hacktron’s security research team has disclosed a two-vulnerability chain that let them compromise multiple OpenAI employees’ ChatGPT accounts — and reach internal OpenAI repositories — in under 72 hours. The findings, published this week, combine a heap overflow remote code execution in the libheif image library (used by OpenAI’s Discourse-hosted community forum) with an SSO misconfiguration in OpenAI’s identity infrastructure.

Because users can connect services like GitHub, Slack, and email to ChatGPT and Codex, the researchers say the theoretical scope was huge. To prove access without learning sensitive information, they opened a harmless proof-of-concept pull request in OpenAI’s internal monorepo. OpenAI fixed the issue roughly 14 hours after the initial submission and paid a $6,500 bounty. The team has since expanded its research into “HEIF Heist,” tracing libheif across Slack, Meta, GitHub Enterprise, and frameworks like Next.js and Astro — warning that any application processing .heic/.heif/.avif images may be affected.

A Chatbot-Fueled Intel Report Nearly Prompted a Military Boarding of a Chinese Ship

CNN reports that a U.S. intelligence report circulated across the military this spring claimed a Chinese ship in the Middle East was transporting components of a nuclear weapons program — and that it was only just before a planned interception that officials discovered the report had been generated with the help of an AI chatbot that inaccurately identified the cargo. The report was, according to one source, “entirely false” but “almost started a war.” Armed U.S. military members were preparing to board the vessel, with military planes in the air, before the error was caught.

The episode underscores the risks of AI-assisted targeting. An analyst queried a chatbot about intelligence on the ship’s manifest, and the bot fused open-source and signals intelligence into a standard-format report that was trusted and disseminated. Senior officials note that AI-tool proliferation across the military is decentralized, with no unified standard for verifying AI-generated information. One source’s blunt warning: “AI allows you to get to a bad idea faster.”

Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint

PrismML has released Ternary Bonsai 2 27B, built on Qwen3.8 27B, claiming 98.2% retention of full-precision benchmark performance in a 5.9GB footprint — more than 9x smaller than its full-precision counterpart. The model uses ternary {−1, 0, +1} weights with FP16 group-wise scaling, achieving 1.76 effective bits per weight, and supports a 262K-token context window with multimodal text-and-image input under the Apache 2.0 license.

PrismML reports the model scores 83.9 across reasoning, math, coding, vision, and agentic tool use, and reaches up to 143 tokens/second on an RTX 5090 and 46.8 tokens/second on an M5 Max. On an RTX 4090 it consumes just 0.714 mWh/token — about 40% more energy-efficient than an 8B model in full precision. The company, founded by Caltech researchers with backing from Khosla Ventures, Cerberus, and Google, argues low-bit models are reshaping the economics of AI deployment from personal devices to data centers: nearly the same capability, in a fraction of the memory and power.

ZCode, the GLM Coding Agent, Is Accused of Silently Uploading Entire Git Histories

A reverse-engineering walkthrough published September 18 by a developer going by “ferstar” alleges that ZCode — the AI coding desktop app from Z.ai, the Beijing-based company behind the GLM open-weight models — silently packages a user’s entire workspace, including complete .git history, and uploads an encrypted archive to Aliyun OSS whenever the app is logged in. The researcher captured a 313MB encrypted archive built from a 345MB commercial workspace spanning 42,411 files.

According to the analysis, ZCode uses envelope encryption: the payload is encrypted with a symmetric key wrapped in an RSA-OAEP public key whose private half lives only in Z.ai’s cloud — meaning the user cannot decrypt it. The developer concludes: “A key that only the server can use serves exactly one purpose: making sure the server can read your code whenever it wants.” UI toggles labeled “Optimize Experience” and “Repo Snapshot Indexing” reportedly only control training authorization and server indexing, not the packaging-and-upload itself, which the analysis says runs unconditionally as a host-level sidecar. Z.ai’s privacy policy, the researcher found, does not mention packaging and uploading entire workspaces and git histories.

The story spread quickly in both English and Chinese, passing hundreds of thousands of views. The most-cited community response came from open-source AI developer Petri Kuittinen: “do NOT trust closed source AI harnesses.”

That’s the week in AI: legal reckoning over how models were built, hard evidence of how fragile their security boundaries can be, fresh lessons about why verification matters in high-stakes deployments, and two reminders — one impressive, one disturbing — of how much power and reach this technology now holds. We’ll be back with more tomorrow.

Friday, September 18, 2026 — This week’s AI news is dominated by open-weight momentum, tooling, and a live glimpse inside one of China’s most transparent model labs. From Nvidia’s first real push into native Rust GPU programming to Xiaomi streaming its post-training metrics in real time, here are the five stories that mattered most.

1. Nvidia moves into native GPU programming in Rust

Nvidia officially leaned into Rust as a first-class language for GPU kernel development, announcing two complementary tracks that together make CUDA Rust a serious option alongside the mature CUDA C++ and CUDA Python toolchains. The first, cuda-oxide, provides a custom rustc codegen backend that compiles SIMT-style GPU kernels written in Rust directly to PTX, built on the Pliron IR framework and LLVM. The second, cutile-rs, brings Tile-based GPU programming to stable Rust, letting the compiler manage thread mapping and memory layout through CUDA Tile IR JIT compilation.

Both projects enforce memory safety at compile time — cuda-oxide through DisjointSlice and launch contracts that prevent aliasing, cutile-rs through tensor partitioning and ownership guarantees. cutile-rs runs on stable Rust 1.89+ with CUDA 13.3, is already published on crates.io, and is in production use inside Hugging Face’s Grout inference engine and mistral.rs, while cuda-oxide remains in early alpha. Nvidia says it plans to support inter-language interoperability across CUDA Rust, C++, and Python, so the choice of frontend won’t lock developers out of the wider ecosystem.

2. Mistral and Mozilla bring private, multilingual AI to the browser

French open-weights lab Mistral and Mozilla announced a partnership to bring open, private, and multilingual AI directly to the web browser. The collaboration pairs Mozilla’s browser distribution and user reach with Mistral’s pragmatic, open-model philosophy, positioning a privacy-respecting, built-in AI assistant as an alternative to cloud-gated incumbents. For Mistral it extends a fast-growing enterprise story — recent partnerships include Cloudera for specialized, sovereign intelligence in enterprise data — while giving Mozilla an on-device AI lane that fits its privacy brand.

3. A 4B model learned to beat Postgres’s own query plans — by 81%

Rohan Bansal trained a small, open-weights Qwen3.8B-Distill model to produce Postgres query plans that run dramatically faster than the database’s default optimizer, posting an 81% speedup on join-heavy workloads. The result: a 44.7% latency reduction across 113 join-heavy queries — notable because the model initially couldn’t produce any plan at all for 99 of them. The project leans on agentic reinforcement learning: Qwen generates candidate strategies per rollout, Postgres measures them against its own default plans, and scalar rewards flow back to nudge the weights toward faster joins.

Join ordering is known to be NP-hard, but verifying whether a chosen plan is good is tractable — execution time gives a single, easy axis to optimize. Bansal’s rig combines SFT and a custom GRPO variant scored in a deliberately noisy environment, splitting RL across a rented 2x H100 node running vLLM and the trainer, with four Postgres containers running on his desk. Off-policy distillation across hundreds of GPT-6 Astra agent trajectories seeded the smaller model, and benchmarks cover the Join Order Benchmark and Cardinality Estimation Benchmark.

4. Xiaomi streams Mimo 2.6’s training live from the trainer’s logs

Xiaomi’s MiMo lab published a live post-training dashboard for its upcoming MiMo 2.6 models, streaming reinforcement-learning metrics in real time for both the mimo-v2.6-pro and mimo-v2.6-flash runs. The page exposes training metrics directly from the trainer’s logs — a level of transparency that Hacker News commenters were quick to contrast with larger US labs. Community observers noted the pro run started September 15 at 10:32 UTC, and estimated the compute burn at roughly $5 per second — about $432K per day, enough to rent 3,000 B300 nodes on Modal. The previous generation, MiMo 2.5, scored 19% on DeepSWE 1.1, and commenters tracking open-weight coding models were watching whether 2.6 closes the gap toward leaders like Fable (70%) and Astra (74%).

5. Anthropic merges Cowork and chat into one Claude

Anthropic is unifying its formerly separate Claude Cowork and chat surfaces into a single Claude, removing the friction of deciding where a task belongs. Questions and long-running work now live in one place, and Claude can continue a handed-off task even after you close your laptop. The change ships with two brand-new products — Claude Docs and Claude Slides — and brings Claude Design directly into conversations, so a report and its leadership deck can come from the same thread, with the slides already matching the doc.

The rollout reaches Pro and Max plans on web, desktop, and mobile over the coming weeks, with Team and Free plans to follow and Enterprise admins given at least 30 days’ notice. Claude defaults to asking before taking actions, though users can let it keep working and check in only when something needs a closer look — keeping humans in the loop on the final say.

That’s the AI news for Friday, September 18, 2026 — from Nvidia’s Rust push to a 4B model that out-plans Postgres, an unusually open week for the field.