Google Unleashes Gemini 4 Argon as Hackers Raid Pentagon Personnel Files

AI & Machine Learning

Google has released Gemini 4 Argon, which it calls its most powerful model yet, aimed at coding, research, writing and cybersecurity work. For now the model is limited to selected cyber partners in Google’s Fairwind Program security initiative, where Google says it can “autonomously find, validate, and patch critical software vulnerabilities.” Google also highlights strong debugging, codebase migration and visual analysis of video and charts, and says Argon tops the Vals AI model index ahead of OpenAI’s GPT-6 Astra and Anthropic’s Fable and Opus models. No pricing was disclosed. The launch is the first Gemini version-number jump in a while and lands as the Gemini app passes one billion monthly users, keeping the frontier race between the three big labs neck and neck. Source: TechCrunch Verified: True

OpenAI has accused Chinese lab Moonshot AI, maker of the Kimi models, of running a coordinated “distillation attack” to extract the reasoning abilities of its models. OpenAI says it logged 16,000 requests from more than 4,000 users on 24 and 25 July alone, with related activity across over 15,000 accounts in a campaign that ran from 1 July until OpenAI shut it down on 28 July. According to the company, the requests manipulated model interactions so protected reasoning could be reproduced in a visible form, in breach of its terms of service. OpenAI banned the accounts, tightened signup controls, closed a pathway that allowed encrypted reasoning to be replayed, and shared its findings with other AI firms and government agencies. The episode sharpens the debate over model distillation as a route for rivals to copy frontier capabilities without the safety guardrails attached. Source: The Register Verified: True

Voice AI company ElevenLabs has doubled its valuation to $22 billion through a $300 million employee tender offer. The secondary sale was co-led by Wellington and T. Rowe Price, both of which intend to keep their stakes through a potential public listing. ElevenLabs, founded in 2022, builds synthetic voice and sound-effect models, and describes the tender as a retention tool to stop staff being poached by competitors. Its previous tender, in September 2025, was $100 million at a $6.6 billion valuation. The jump shows investors are still paying up for specialist model companies even as attention concentrates on the general-purpose labs. Source: TechCrunch Verified: True

Consumer Hardware

Reddit is shutting off RSS feeds on 13 November 2026 and will end free public API access in March 2027, citing large-scale scraping by AI bots. AI assistants and researchers that want Reddit content will need to negotiate commercial data deals instead. Old Reddit will be limited to logged-in users who have used it within the past six months, and third-party app developers must register with Reddit by 12 January 2027 or lose API access. The moves track Reddit’s growing data-licensing business, with “other revenue” up 24% year on year to $43 million. For users, it marks the end of a long-standing way to follow subreddits outside Reddit’s own apps. Source: TechCrunch Verified: True

DoorDash has unveiled an AI agent that lets customers order food by texting through Apple Messages. Users can send requests like “order my usual,” and the agent remembers preferences, recommends nearby restaurants with photos, and coordinates group orders with mixed dietary needs. The feature is in a waitlist phase for US users with no firm launch date yet. DoorDash also announced drone delivery tests with selected Northern California restaurants. The move pits DoorDash against Uber Eats and Grubhub on hands-free ordering as personal AI agents become a new storefront for consumer services. Source: TechCrunch Verified: True

Cybersecurity

Hackers spent months inside the US Defense Manpower Data Center (DMDC), stealing records on about 2.8 million living people and 300,000 deceased individuals between October 2025 and mid-July 2026. The attackers exploited a flaw in an unnamed file-sharing system to take unencrypted personal data including Social Security numbers, names, dates of birth and military service records. DMDC holds more than 60 million records and acts as the identity backbone for Pentagon access credentials, which makes the haul especially sensitive. The Department of Defense says it has “no indication that the information was misused,” and the attackers have not been identified. It follows a string of intrusions at federal agencies, including the FBI earlier in September. Source: TechCrunch Verified: True

Cisco has warned that a critical authentication bypass in Catalyst SD-WAN Manager, tracked as CVE-2026-76504, is being exploited in the wild. The flaw stems from improper handling of URI encoding, letting attackers send crafted HTTP requests (using “%6a” in place of the letter “j”) to gain unauthenticated admin access regardless of configuration. A single SD-WAN Manager can control up to 6,000 network devices, so a compromise gives attackers broad reach into corporate networks. Cisco has shipped fixes for every affected branch, and CISA added the bug to its Known Exploited Vulnerabilities catalog with a 3 October patch deadline for federal agencies. It is the latest in a run of edge-device zero-days hitting network management planes this autumn. Source: BleepingComputer Verified: True

A 16-year-old researcher found that Microsoft’s internal Titan analytics platform failed to verify JWT signatures, allowing admin access simply by editing an unsigned token’s username to “admin.” The flaw exposed potential access to databases holding 17.3 trillion rows, including employee records, dashboards and analytics configurations. Titan is meant to be restricted to Microsoft employees. Microsoft fixed the issue after coordinated disclosure and paid a $5,000 bug bounty on 17 September. The case is a reminder that a single missing check in token validation can undo every other layer of internal access control. Source: The Register Verified: True

Enterprise Infrastructure

A Microsoft infrastructure servicing operation disrupted Azure networking across 18 regions, including West US, North Europe, West Europe, Southeast Asia and Japan West. ExpressRoute Gateway, VPN Gateway, Azure VMware Service, Azure Firewall, Application Gateway and Web Application Firewall all saw degraded or interrupted connectivity. Microsoft said it found a correlation with operating-system servicing activity, paused that work, and moved to restore network management components using healthy instances while it investigates the root cause. Hybrid cloud customers that link on-premises sites to Azure bore the brunt, though some gateways only lost redundancy rather than going fully offline. The incident underlines how routine maintenance on shared control planes can ripple into customers’ private networks. Source: The Register Verified: True

Micron says the memory shortage will get worse through 2028, with its CEO warning customers to expect “much higher prices” than today. The company reported annual revenue of $133.2 billion, up from $37.4 billion, and net income up 895% to $85 billion, driven by AI demand. Gross margin in its cloud memory unit, which sells high-bandwidth memory, rose to 83% from 59% a year earlier. Micron is working with Nvidia on custom HBM4E for next-generation GPUs, plans $25 billion in capital spending, and guided to $61.5 billion in revenue for the first quarter of fiscal 2027. For buyers of servers and PCs, the outlook means memory will remain one of the most constrained and expensive parts of the bill of materials. Source: The Register Verified: True

Huawei rotating chairman Eric Xu claims the company’s Ascend NPUs now outsell Nvidia’s AI chips in China. Xu conceded Ascend trails rivals technologically but argued that guaranteed supply, free from sudden US export bans, is winning customers, helped by Beijing pushing domestic operators toward local silicon after Washington’s shifting export rules. Huawei is building large computing clusters and an architecture designed to scale to one million NPUs, though it says it lacks capacity to serve markets beyond a few friendly countries. Better software tooling that reduces dependence on Nvidia’s CUDA is making the switch easier for developers such as DeepSeek. If the claim holds, it marks a turning point in China’s effort to build an AI infrastructure stack independent of US suppliers. Source: The Register Verified: True

Policy & Regulation

The Third Circuit Court of Appeals has ruled that ROSS Intelligence’s use of Thomson Reuters’ Westlaw headnotes to train a competing AI legal research tool was not fair use. Judge Tamika Montgomery-Reeves wrote that the headnotes carried enough “creative spark” to be protected and that ROSS’s use was “minimally transformative at best” because it served the same commercial purpose as the original. The decision on 30 September affirms the 2025 district court ruling by Judge Stephanos Bibas. It is the first time a US appeals court has weighed in on fair use for AI training. While the case involves a non-generative system, it hands copyright holders a significant precedent as dozens of lawsuits against generative AI companies move through the courts. Source: Courthouse News Service Verified: True

Defense Secretary Pete Hegseth has launched Project Meridian, a 120-day study of future warfare co-led by Elon Musk, Anduril founder Palmer Luckey and former House Speaker Newt Gingrich. The group, which also includes academics and policy experts, will work under Defense Department CTO Emil Michael to “study the battlefields of the future” and recommend capabilities the US military should develop. Both SpaceX and Anduril already hold Pentagon contracts for satellites and autonomous systems. Critics warn of conflicts of interest, since the study’s recommendations could steer spending toward the companies its leaders run. The initiative gives tech executives an unusually direct hand in shaping US defence technology priorities. Source: TechCrunch Verified: True