AI hallucination nearly sent the U.S. military to board a Chinese ship

The U.S. military was preparing to intercept and board a Chinese ship in the Middle East, with air support, after a Special Operations Command analyst’s report claimed the vessel was carrying nuclear-arms-program components. Officials later found the material identification was entirely false, the result of a chatbot used to analyze the ship’s manifest fusing open-source intelligence with secret signals intelligence; one source familiar with the episode said it almost started a war.

The account, reported by CNN and attributed to four sources familiar with the episode, was not independently reviewed beyond that reporting, so the precise tool and chain of events remain single-sourced. It lands as the Pentagon accelerates military AI use, including a January AI acceleration strategy pushing data across services for AI exploitation and a bespoke GenAI.mil platform built on Google’s Gemini for Government.

Featured source: Ars Technica .

Other Stories

New unsealed filing says OpenAI and Microsoft warned of an AI content “doom loop”

Newly unsealed court documents in the New York Times’ case against OpenAI and Microsoft quote internal material warning that the companies’ AI content strategy started a “doom loop” harming both the web and model performance, alongside Microsoft’s Brent Hecht calling large-scale data harvesting the “largest theft of labor in human history.” The 92-page filing also quotes OpenAI’s Greg Brockman on potential “gazillions” in commercial returns and cites examples of GPT-4 reproducing passages from publishers including the Times, the Mercury News, and the Denver Post.

These are quotes as presented in the Times’ adversarial filing, not adjudicated findings, and Microsoft says Hecht’s comments are one employee’s individual perspective rather than company views. A related filing in the same suit alleged paywall bypassing and large click-through declines; the fair-use question at the case’s core remains unsettled.

Filed from: The Verge .

Manus seeks $500 million at a $4 billion valuation as an independent company

Chinese AI startup Manus is in discussions to raise $500 million at a $4 billion valuation after resuming operations as an independent company, The Wall Street Journal reported, citing anonymous sources. Named potential participants include IDG Capital, Boyu Capital, battery maker CATL, and existing backers Tencent, HSG, and ZhenFund, with a possible restructuring toward a Hong Kong IPO also said to be under consideration.

The figures and investor names are sourced anonymously and unconfirmed by Manus, which did not immediately comment. The backdrop is established: a $2 billion Meta acquisition announced last December, blocked by Beijing over export-control and foreign-investment concerns, then unwound with early backers helping buy back shares at roughly $2 billion, followed by a data-deletion notice to users in August tied to jurisdictional requirements.

Filed from: TechCrunch , Wall Street Journal .

Google turns its CC agent toward running family households

Google is testing a family-focused version of CC, its AI agent that works across email, calendar, chats, and tasks. The agent gets its own Google account so up to six family members can each choose what to share, and it can track school and activity emails, add dates to a shared calendar and task list, fill out permission slips and registration PDFs, build shopping lists and meal plans, and create shared Docs and Sheets, asking for missing details as needed.

Google says CC runs in an isolated environment with its own permissions; the rollout’s data-handling behavior comes from Google’s announcement as reported, and independent testing of the new household features is not yet available.

Filed from: TechCrunch , Google on CC .

FAA readies a $875 million AI tool for DC airspace

The FAA is preparing to switch on SMART, an AI system that predicts traffic flows and conflicts from schedules, weather, airport capacity, and airspace conditions, starting with the three major Washington, DC-area airports. Government and industry officials put the debut as soon as Monday, September 21, ahead of a planned nationwide rollout, inside a $875 million, 12-year contract awarded in June to Boston-based Air Space Intelligence.

Airlines were reportedly confused for weeks about deployment plans, with concerns easing only after the FAA said SMART would not change controller or airline procedures and would surface alternative route information through existing systems. What model types SMART uses is unclear, and one former FAA official wants written evidence gates, including real-workload performance and clear ownership when a prediction is wrong, before wider expansion.

Filed from: Ars Technica .

Newsom orders AI safety recommendations and floats a “kill switch”

California Gov. Gavin Newsom signed an executive order directing the state to convene experts to deliver recommendations within two months on strengthening AI safety law. The order asks the group to consider onsite independent verification, audited transparency reports and risk assessments, a routinely verified “kill switch” for frontier models, and mandatory reporting of loss-of-control incidents.

The order directs work rather than imposing mandates itself, and it also speeds implementation of two recently signed laws on independent verifiers and an AI auditor registry. Newsom framed the framework as a model for federal action, calling on Congress and the administration to adopt it or use it as a floor.

Filed from: The Verge .

Computer science graduates face recession-like job prospects as AI exposure bites

Three U.S. Census Bureau economists find that graduates in the most AI-exposed majors saw their likelihood of initial employment fall by five percentage points and full-quarter initial earnings fall by 13 percent after ChatGPT’s late-2022 arrival, a decline the authors say rivals graduating into a large recession. About half of the earnings drop comes from graduates shifting into lower-paying industries such as retail and food service, with computer science, mathematics, statistics, some engineering disciplines, networking, and journalism among the most exposed fields.

The paper is a working paper rather than a peer-reviewed finding, and the authors acknowledge alternatives such as self-employment and graduate-school enrollment could account for as much as half of the employment decline. Effects ease with time since graduation, though prior recession research suggests early-career weakness can leave lasting earnings scars.

Filed from: The Register , Census Bureau working paper CES-WP-26-56 .

India orders caller-ID apps to share spam reports with telcos

India’s telecom regulator TRAI amended commercial-communications rules to require caller-ID and call-management apps that let users flag spam to forward those reports to a blockchain-based platform run by telecom operators. Truecaller, whose largest market is India with well over 350 million of its 500 million-plus monthly users, called the one-way sharing anti-competitive and commercially sensitive.

The amendments also retain a restriction barring call-management apps from blanket blocking, filtering, or spam-tagging calls from designated number series used for promotional, service, and transactional communications, though individual users can still block such calls on their own devices. The rules land at large scale: Truecaller says its Indian users encountered around 42 billion spam calls in 2025.

Filed from: TechCrunch .

From the Community

ZCode client silently uploaded full Git history to Aliyun OSS, researcher finds

A developer reverse-engineering Zhipu’s ZCode AI coding desktop app reports that, whenever a user is logged in, the client packaged the entire workspace, including complete Git history, LFS caches, reflogs, and global app configs, encrypted it, and posted it directly to Aliyun OSS. The write-up describes a 313MB encrypted archive left in a local pending directory after hundreds of failed upload attempts, with Git data accounting for the large majority of a tens-of-thousands-of-files snapshot, and says UI toggles did not stop the capture while the content key was wrapped with a server-supplied RSA public key.

The account is a single-author reverse-engineering investigation, AI-translated by its author, with detailed local evidence; the vendor’s response, appended to the post in a September 19 update, was not independently reviewed here.

Filed from: Inside ZCode .

OpenAI says its own LLMs cut Jalapeño accelerator design to under 20 months

IEEE Spectrum reports that OpenAI’s Jalapeño inference accelerator moved from first architecture concept to first silicon in under 20 months, with nine months between first RTL and tape-out, described as delivering up to 13.4 petaflops of 4-bit compute with 232GB of HBM4. Hardware VP Richard Ho says the design team averaged fewer than 100 people, with Broadcom handling physical design from the gates onward, the front-end workflow built around Google’s open-source XLS high-level synthesis chain, and AI-guided physical design cutting matrix-multiply unit area by 10% against an optimized human baseline.

Performance and timeline figures are largely vendor-reported, with independent experts quoted calling the schedule credible while crediting Broadcom’s implementation role as essential.

Filed from: IEEE Spectrum .

NASA and IBM release an open-source lunar foundation model

NASA and IBM Research released a lunar foundation model pretrained from scratch on SomBench, a multimodal dataset of nearly two million co-registered data bundles spanning 11 modalities and two spatial scales, from imagery and topography to mineralogy, radar, and gravity. USRA contributed planetary science expertise and dataset development, including a high-resolution crater benchmark with more than 49,000 manually identified craters. Across crater detection, irregular mare patch segmentation, and lunar polar ice prospectivity, the pretrained model matched or outperformed ImageNet-pretrained and non-pretrained baselines, with the strongest gains in label efficiency for crater detection.

The performance figures come from the project’s own benchmark suite, with independent evaluation not yet available, and the model, fine-tuning code, and datasets are released openly through Hugging Face.

Filed from: USRA .