Research finds Blackwell fingerprints in six models from China's leading AI labs
A forensic report says DeepSeek, Moonshot, MiniMax and Ant Group used export-restricted Nvidia Blackwell chips for recent models, while Nvidia's board looks outside the company for a chief executive.

On this page
Six models released since April carry the numerical fingerprints of Nvidia Blackwell accelerators, according to a forensic report published on 12 November. The models come from DeepSeek, Moonshot AI, MiniMax and Ant Group. Unemployment Labs, a nonprofit research group led by former OpenAI researcher Will DePue, produced the analysis. Earlier models from the same four labs match Nvidia’s older Hopper generation instead.
The report works from published model weights rather than leaks or insider accounts. Its author recovered traces of each model’s initial values and searched for hardware configurations that reproduce them. For two models, DeepSeek-V4-Flash and Kimi K3, he confirmed the match on physical Blackwell chips. That shows how the weights were created and supports the inference that Blackwell also ran the training, without proving it. The report does not establish where the chips were located, how access was obtained, or whether export controls were violated. It treats specially configured AMD hardware as the only alternative it could not fully rule out.
Blackwell parts require a US licence for export to China, and no such licence has been reported. Nvidia’s former chief executive, Jensen Huang, resigned this week after admitting an affair with an executive at Megaspeed. The Singapore customer has been investigated by US officials over chip shipments.
Featured source: Unemployment Labs , TechCrunch .
Other Stories
OpenAI’s ChatGPT Wallet lets agents buy things with a tokenized card
ChatGPT Wallet began rolling out on 13 November to ChatGPT Pro subscribers in the United States. Users must be at least 18 and pay with an eligible US-issued Visa, Mastercard, American Express or Chase card. Dots agents can then buy things on a user’s instruction, but each purchase needs approval and a biometric check such as Face ID or a fingerprint. OpenAI says it swaps the card number for a payment token, as phone wallets do, so neither the agent nor the merchant sees the underlying details. The company expects users eventually to delegate purchases within limits they set, and notes that some retailers block AI agents.
Filed from: OpenAI , The Verge .
Nvidia will have to look outside for its next chief executive
Nvidia’s search for a permanent chief executive cannot draw on an internal heir. Jensen Huang named no deputy, and about 60 employees reported directly to him, The Verge reported on 14 November. Analysts told the publication that the product plan is secure for now. Moor Insights’ Patrick Moorhead said the roadmap and manufacturing are locked in through 2027. Creative Strategies’ Ben Bajarin said the next generations are already in design, in tape-out or on order. Jon Peddie said nobody inside the company wants the job. The analysts called customers hedging during a leadership vacuum a bigger risk than any rival chip.
Filed from: The Verge , Rest of World .
Attackers used 4,837 GitHub accounts to plant credential-stealing workflows
Security firm Aikido reported a campaign it tracks as wfx-loot. Attackers used 4,837 compromised GitHub accounts to open pull requests against 61,764 repositories. The requests were titled “docs: dependency audit notes” and came from branches named wfx-docs, working as probes to see which projects those accounts could modify. More than 10,000 reached repositories the accounts did not own. Where the attackers had write access, they chose one of three methods. They edited existing GitHub Actions workflows, added a shell script triggered by an npm install hook, or created a workflow on a temporary branch. Aikido says 1,835 repositories ran the payload, which hunted for cloud credentials, SSH keys, API keys and shell histories. The researchers did not say how the accounts were compromised.
Filed from: TechRadar Pro , BleepingComputer .
FAA extends its AI flight-advisory system to all US air route centres
The Federal Aviation Administration has expanded an AI advisory system from the airspace around Washington, DC, to all 22 air route traffic control centres in the United States. The agency values the system, called SMART, at $875 million. It says SMART rerouted 180 flights and saved roughly 10 minutes and almost 80 gallons of fuel per flight after going live on 21 September. A team of FAA specialists in Washington validates its recommendations, and local controllers may decline them. The union representing more than 15,000 controllers told Flying it was not involved in the design, testing or implementation.
Filed from: Ars Technica , Department of Transportation .
Cloudflare publishes the source code for Kitesurf, its browser for AI agents
Cloudflare released Kitesurf under the Apache 2.0 licence, so developers can run the browser locally or deploy it to their own Workers accounts. Kitesurf parses HTML, applies CSS, executes page JavaScript and captures screenshots without a browser container. It speaks the Chrome DevTools Protocol, which lets clients such as Puppeteer attach. The engine is built from Rust packages compiled to WebAssembly, including Blitz for layout and Boa as a JavaScript fallback. A self-hosted deployment needs a Workers Paid plan, because each page runs in an isolated Dynamic Worker.
Filed from: Cloudflare , GitHub repository .
From the Community
Citizen Lab ties an ad-data surveillance system to agencies in more countries
Citizen Lab says Penlink’s Webloc has been bought by security agencies beyond the customers its April report identified in the US, El Salvador and Hungary. The system builds surveillance on advertising data. A Bloomberg investigation named the UK, Netherlands, Germany, Mexico, Singapore, Saudi Arabia, the UAE and Israel. Citizen Lab says that list matches potential customers it had already flagged. The group notes that the lawfulness of ad-based mass surveillance is contested in Europe and the UK.
Filed from: Citizen Lab .
Anthropic gives Claude Code Projects a coordinator that reports unprompted
Anthropic has updated Claude Code Projects so the project coordinator posts thread status and answers questions in the main chat without being polled. Threads can now be shared through read-only links with signed-in Claude users, except those that touched a local computer. The update adds /btw and /fork commands, scheduled reminders and artifact thumbnails. Projects remains a public beta for Pro and Max subscribers. Every parallel worker draws on the same Claude Code allowance, and accounts are capped at 200 new threads a day. The feature details rest on a secondary report, and Anthropic’s own release notes were not reviewed.
Filed from: AlphaSignal .
ggml reaches 1.0.0 and prunes its deprecated interface
The tensor library behind llama.cpp tagged version 1.0.0. It removes the deprecated public API and the legacy DIAG_MASK operations, and bumps the RPC protocol to v9 so incompatible peers are rejected. The release adds configurable YaRN truncation and FP8 support. It also carries performance work across backends. That includes few-row quantized matmuls and mixture-of-experts fusion on CUDA and Metal, a Vulkan prefill FlashAttention path, a device-side MoE scheduler on SYCL and new WebGPU kernels. The notes describe a sync of upstream llama.cpp changes, and the speedups have not been independently benchmarked.
Filed from: ggml v1.0.0 release .



