Tech Express | October 6, 2026 — US Pentagon formally stops using Anthropic's AI tools

1. Amid the AI arms race, the US Pentagon formally stops using Anthropic's AI tools

Source: BBC

The US Department of Defense has stopped using Anthropic's artificial intelligence tools, an official told the BBC. In a statement on Monday (October 5), the defense official said the Pentagon "has ceased using Anthropic's products." Months earlier, the Pentagon had designated the company a supply chain risk on national security grounds.

Defense Secretary Hegseth labeled Anthropic a supply chain risk in February and said the Pentagon would phase out its products by late August; it is still unclear why the delay occurred. Multiple people familiar with the matter told the BBC that until last week, the department was still using Anthropic's Claude model, including its powerful Mythos model, for research, analysis, intelligence gathering and military operations against Iran.

Anthropic refused to remove Claude's safety guardrails, citing concerns over mass surveillance and autonomous weapons, after which Hegseth designated the company a "national security supply chain risk." Anthropic called the move "unprecedented and unlawful" and has sued the Trump administration. Lauren Kahn, a senior research analyst at Georgetown University's Center for Security and Emerging Technology, said that because Claude is deeply embedded in the Maven Smart System run by Palantir, removing it from the system quickly is not easy.

Source

2. 'US version of DeepSeek' Reflection launches Beam model, still a generation behind V4.1 Flash

Source: Mydrivers

In the open-source AI model field, US company Reflection, described as the "US version of DeepSeek," has officially released its Beam large model. Beam is a 501-billion-parameter MoE model that activates only 23 billion parameters, supports a 1-million-token context window and is a text-only model. It was trained for four weeks on 10,500 Nvidia B300 GPUs with 23.8 trillion tokens of pre-training data, and will be fully open-sourced under the Apache 2.0 license this month.

On performance, Beam was compared with several Chinese open-source models as well as Nvidia's earlier Nemotron series, and it is much stronger than previous US open-source models. But compared with Chinese open-source AI, Beam only matches the level of GLM-5.2, still about a generation behind DeepSeek V4.1 Flash, and it also loses to Chinese models such as K3, GLM-5.3 and Qwen 3.8 Max.

Beam's specific price has not been announced, but the company says its compute requirement is only one-quarter to one-third that of GLM-5.2, suggesting a clear cost advantage. After the launch, Reflection's CEO leaned on a narrative of Western anxiety, stressing that if the West does not build open-source AI, global AI will be built by Chinese models.

Source

3. Study: Windows malware uses Grok AI to stay hidden

Source: Fox News

Security researchers have uncovered a new piece of Windows malware, called x47.c, that gives cybercriminals many ways to cause trouble from a single infected PC. It can steal passwords, grab browser cookies, route internet traffic through the victim's computer and even burn through paid AI credits. According to Fox News, x47.c can reportedly use xAI's Grok to help decide how to keep itself running on an infected Windows machine.

The malware was uncovered by researchers at Qrator Research Labs while tracking cybercrime activity. A threat actor using the name WraithTools has been advertising access to the malware, which comes with tools for stealing credentials and launching attacks. Qrator based its findings on the seller's advertisement, technical documentation, screenshots and follow-up messages, so the research shows what x47.c is advertised and designed to do rather than how widely it is currently infecting Windows PCs.

Qrator found 18 advertised attack methods built into x47.c. One targets paid AI accounts: if an attacker already has a valid API key, the malware can repeatedly send requests to an AI provider, using up prepaid credits or raising the victim's bill, which Qrator describes as a "Denial of Wallet" attack. The seller also advertises an "AI Stealth" feature that can use Grok to look at the state of the infected computer and select from a predefined list of ways to maintain access.

Source

4. YMTC expects NAND flash shortage to last about three more years

Source: Mydrivers

Chinese memory maker YMTC expects the current global NAND flash supply crunch to last about three more years. If that forecast holds, the supply-demand imbalance in the NAND market could extend to around 2029, and prices for the memory chips used in SSDs, phones and other consumer electronics could remain elevated for a long time. The view came from YMTC's recent discussions with investors and market participants.

As AI infrastructure continues to expand, data centers' demand for high-capacity enterprise SSDs is growing rapidly, and more and more NAND capacity is being used for servers and AI-related storage, squeezing the supply available to the traditional consumer electronics market. The current crunch is not limited to NAND: DRAM is also severely undersupplied, and the high-bandwidth memory (HBM) used by AI accelerators is consuming more and more wafer capacity.

YMTC itself is also expanding its NAND capacity sharply. Its third wafer fab in Wuhan is expected to start operating by the end of 2026 and gradually ramp up output in 2027. It has been reported that YMTC even plans to become the world's largest NAND flash manufacturer by the end of 2027. Samsung still ranks first in the global NAND market, followed by SK Hynix and its subsidiary Solidigm.

Source

5. Up to 192GB unified memory: AMD unveils Ryzen AI Max Pro 495

Source: Mydrivers

Nvidia and Microsoft are set to hold an RTX Spark launch event on October 7, and AMD has gotten ahead of it by two days, showing off the Ryzen AI Max Pro 495. The chip, codenamed Strix Halo, supports up to 192GB of unified memory, packs 16 desktop-class Zen 5 cores, integrates an RDNA 3.5 graphics engine and offers a memory rate of 8.5GT/s.

The 192GB of unified memory is one of its biggest highlights. Using AMD's variable graphics memory (VGM) slider, users can allocate up to 160GB of memory to graphics processing, letting a single machine fully run a 100-billion-parameter-class large model offline, with no cloud subscription needed and no private data such as tax files leaving the machine.

Nvidia's RTX Spark uses an Arm CPU paired with a Blackwell-architecture GPU and also pushes unified memory, but the Arm approach cannot avoid Microsoft's Prism emulation layer in the Windows ecosystem, which, despite improvements, still brings power loss and compatibility issues. AMD, by contrast, repeatedly stresses "native x86, no emulation layer," running directly on both Windows and Linux.

Source

6. Apple opens iPhone Duo app submissions ahead of October 23 launch

Source: MacRumors

Apple today announced that apps optimized for the iPhone Duo can now be submitted for App Store approval, ahead of the foldable device launching on Friday, October 23. While existing App Store apps will automatically run on the iPhone Duo, only apps built with the iOS 27.1 SDK or later can dynamically resize to fill the entire inner display, whereas unoptimized apps will show black borders.

Developers can get their apps ready for the iPhone Duo using Xcode 27.1, and Apple shared some tips on its website. According to MacRumors, Apple's first foldable iPhone has a book-style design with a 5.4-inch outer display and a 7.6-inch inner display with a minimal crease down the middle.

Apple also lists other upcoming products, including a squarish 7-inch screen for controlling smart home devices, listening to music and making video calls, plus new models with faster chips and new colors.

Source

7. Apple's touchscreen OLED MacBook Pro may debut next month

Source: Mydrivers

Journalist Mark Gurman reports that Apple's redesigned MacBook Pro is expected to launch next month. The update is said to include a touchscreen OLED display and to make the chassis "noticeably lighter" by shrinking internal parts and rearranging the component layout. The new machines will come in 14-inch and 16-inch sizes.

Apple's current equivalent models use M5 Pro or M5 Max chips; the 14-inch weighs about 1.60 to 1.62 kg and the 16-inch about 2.14 to 2.15 kg. These are current-generation figures and cannot be used to estimate how much lighter the new models will be. Earlier reports also said the new machine may adopt a thinner body, replace the notch with Dynamic Island and is testing macOS 27.2, but those details have not been confirmed by Apple.

Chip configurations differ across reports. GSMArena says the new models may use the M6 and M5 Ultra, while other reports citing Gurman mention the M5 Pro and M5 Max. Gurman previously said the new MacBook Pro might launch in October or November, with the latest report pointing to next month, that is, November. However, the exact launch date, price, touch interaction method and full specifications have not been announced.

Source

8. After 25 years, physicists can finally measure the 'other' kind of quantum entanglement

Source: ScienceAlert

Quantum entanglement is one of the most cognitively jarring disconnects between the quantum domain and the macroscopic world. When fundamental particles such as electrons or photons are entangled, they share a quantum identity and cannot be described as individual entities, existing in a superposition with no fixed state until at least one is measured. Physicists from Kyoto University and Hiroshima University have pulled us a few steps closer to our quantum future in a landmark study that determined the configuration of multi-particle entangled states, starting with three photons.

Previously, scientists measured Greenberger-Horne-Zeilinger (GHZ) states, the well-known and fundamental type of multi-particle entanglement. In their paper, published in Science Advances, the researchers reported a new technique to measure so-called W-state entanglement types. Quantum information researcher Shigeki Takeuchi said that more than 25 years after the initial proposal concerning entangled measurement for GHZ states, they have finally obtained entangled measurement for the W state as well, with a genuine experimental demonstration for 3-photon W states.

W states are more complex and have an interesting property that differs from GHZ states: in W states, if you lose one of the entangled particles, the remaining ones retain their useful entangled state. The team used a discrete Fourier transform (DFT) optical circuit device of their own making to determine the entire entangled system in a single one-shot measurement. They reported an averaged measurement discrimination fidelity (MDF) of 0.871 ± 0.039, correctly identifying the W state 87 percent of the time and comfortably exceeding the mathematical threshold of 66.7 percent.

Source

9. US-China RISC-V cooperation deepens as open architecture takes 25% processor share

Source: Mydrivers

Cooperation between China and the United States on the open-source RISC-V chip architecture is accelerating, according to reports. MIPS chief technology officer Yankin Tanurhan said the US chip IP supplier is working closely with Chinese companies to jointly advance RISC-V standardization work.

The trend is backed by solid data. According to RISC-V International's public directory, the foundation's institutional members come from 37 countries, with China and the US the two most engaged, holding 89 and 87 members respectively, together accounting for 42 percent of total institutional members. Among the highest-level premier members, China accounts for 50 percent and the US 38 percent. Chinese Academy of Engineering academician Ni Guangnan disclosed at the 2026 XuanTie RISC-V Ecosystem Conference that RISC-V has officially taken 25 percent of the global processor market.

Chinese companies are also accelerating at the technical level. Alibaba Damo Academy's XuanTie C950 uses 8-instruction decoding, a 16-stage pipeline and an out-of-order window of more than 1,000 instructions, with a maximum clock of 3.2GHz; its single-core general performance exceeded 70 points on the SPECint2006 benchmark for the first time. On standards, on August 31, 2026, the Supervisor Physical Memory Protection (SPMP) instruction-set standard led by a Shanghai Jiao Tong University team formally passed the RISC-V International board review and was approved for inclusion in the RISC-V instruction set architecture standard.

Source

10. Boss spends 900,000 yuan on four robot dogs as maker says dispute resolved

Source: Mydrivers

A topic about a boss who spent nearly 900,000 yuan on four robot dogs and was unhappy with their performance has drawn attention recently. Today, the manufacturer involved, Juwei Technology, issued a statement on the recent online discussion, saying the product procurement matter with the customer has now been properly resolved through negotiation. Juwei Technology also said that, for projects involving specific regions and sensitive application scenarios, cooperation terms and responsibility boundaries must be carefully defined based on final use, the degree of technical openness and relevant compliance requirements.

According to a prior report by "1818 Golden Eye," Xiao, the head of Zhejiang Fazhiqing Robotics, ordered two P2 wheeled quadruped robots, or "big dogs," from Hangzhou Juwei Technology on July 29 to carry out an inspection project, for a contract totaling 580,000 yuan with delivery promised before August 10. Xiao paid 70 percent of the price, or 406,000 yuan, but the goods were repeatedly delayed, with the company citing a need to upgrade the robot dogs' motors. Needing the dogs urgently, Xiao ordered two more "Curious S01" medium robot dogs on August 28 at 160,000 yuan each.

One of the "medium dogs" was sent to the US for oil-field inspections, while the other, unpacked in Wenzhou for a demonstration, was found to stand with two unbalanced legs; manufacturer staff could not fix it after an hour of on-site adjustment and said it needed a week of factory repair. More critically, neither medium dog could be used for commercial secondary development, as the devices did not grant low-level development permissions and the system was locked. Xiao said the dogs were purchased to be fitted with gimbals, cameras and a legal AI model for park patrols and home security, and then sold, which counts as commercial use. With both sides at odds, market regulators' mediation failed and they were advised to resolve the matter through legal channels.

Source