AI Scale, Security Crises, and Robot Ambitions Define This Week in Tech

This week’s technology headlines highlighted AI’s expanding reach into consumer products, scientific research, weather forecasting, journalism, robotics, and critical infrastructure. At the same time, costly mistakes, unresolved security risks, and escalating energy demands showed why rapid deployment must be matched by stronger oversight.

Top news

AI adoption accelerates across platforms and devices

Google’s Gemini has surpassed 1 billion monthly users, making it Alphabet’s 14th product to reach that benchmark — and the fastest in the company’s history to do so.

Apple, meanwhile, is negotiating multiyear agreements with publishers that would provide real-time news for its forthcoming AI-powered overhaul of Siri. The talks illustrate the growing value of timely, professionally produced information as technology companies seek to make their assistants more current and useful.

OpenAI is reportedly preparing to expand beyond software with a portable, screenless AI smart speaker priced between $300 and $400. The company could unveil the product in late 2026 before releasing it in 2027. Codesigned with Jony Ive’s LoveFrom, the device would reportedly combine ChatGPT Voice Mode with cameras, microphones, sensors, and mechanical components.

Frontier models, chips, and infrastructure scale up

ByteDance is reportedly training an AI model with as many as 10 trillion parameters, signaling the accelerating scale of China’s efforts to compete at the frontier of artificial intelligence.

Microsoft is also preparing a major infrastructure push. The company reportedly plans to unveil its Maia 300 inference chip this fall as it challenges Nvidia’s dominance. Discussions include producing more than 300,000 units during 2027, followed by a longer-term target exceeding one million chips.

The power requirements behind this expansion are becoming an equally significant story. Amazon is backing a proposed 7.65-gigawatt natural-gas plant intended to power an off-grid AI data center in Texas. The project could produce as much as 33 million tons of carbon dioxide annually, underscoring the environmental cost of rapidly scaling AI infrastructure.

AI delivers scientific and forecasting breakthroughs

China’s Fengwu AI model predicted Typhoon Dolphin’s mainland landfall five days in advance, missing the eventual location by approximately 30 kilometers and the timing by about 30 minutes. Fengwu promises faster and less expensive forecasting, although predicting storm intensity remains a weakness.

In biotechnology, Stanford and Arc Institute researchers used Evo genomic models to design complete viral genomes from scratch. Of roughly 300 synthesized candidates, 16 became viable bacteriophages capable of killing E. coli in laboratory tests. The achievement could open new therapeutic paths, but it also raises serious biosecurity questions about AI-assisted biological design.

Failures and provenance questions complicate AI deployment

The hazards of trusting unverified AI output became painfully clear after a farmer in China reportedly lost nearly 25 acres of sesame after applying an AI-recommended chemical treatment. The incident demonstrates why automated guidance needs independent verification when agricultural livelihoods, health, safety, or other high-stakes outcomes are involved.

Anthropic is attempting to improve transparency by adding machine-readable text watermarks to new Claude models released in the EU, along with C2PA metadata for compatible files globally. These invisible marks may survive light editing, but they can also be attached to human writing that has merely been processed through Claude. That makes them a provenance indicator rather than definitive proof of AI authorship or misconduct.

Robots and app stores challenge established models

Agility Robotics says its forthcoming Digit v5 humanoid robot can work for more than 20 hours per day with the help of rapid charging. Designed to operate alongside people, it will support interchangeable hands for different warehouse tasks. Early units are expected to arrive in December 2026.

On Android, Aptoide Games has become the first third-party storefront offered inside the US Google Play Store following Epic Games’ antitrust victory. Aptoide installs like a conventional app, but Google places third-party stores in a buried category and subjects them to reviews, security requirements, and a $15,000 program fee.

Insider intel

AI newsrooms turn publishing speed into a competitive weapon

A lean automated operation has offered a glimpse of how AI could alter breaking-news competition. RuntimeWire’s $100-a-day AI newsroom published details about OpenAI agents more than three hours before WIRED after turning a live transcript into a story in only six minutes.

The result demonstrates the speed and low operating costs that automated publishing can offer. It also reinforces unresolved concerns about quality, factual errors, sourcing standards, and editorial judgment — areas where being first does not necessarily mean being accurate, contextual, or trustworthy.

Security alerts

Browser extensions and phishing campaigns target credentials

A large brand-spoofing operation used 737 Chrome VPN and proxy extensions to hijack browser traffic. Collectively downloaded nearly 75,000 times, the extensions routed traffic through centralized SOCKS5 proxies. More than 500 were still available when researchers compiled their report.

Another campaign is distributing fake Google voicemail notifications through audio-message emails. Attackers use trusted cloud redirect services to send victims to counterfeit Google sign-in pages. Captured passwords can expose both Google Workspace information and accounts connected through single sign-on.

Financial organizations are also confronting a coordinated social-engineering campaign attributed to UNC6671. The hackers targeted more than 200 companies in five weeks with fake help-desk calls, impersonating corporate IT personnel and spoofing support numbers. Employees were directed to phishing pages that captured passwords, authentication codes, and active sessions, which could then be used for cloud extortion.

Windows and Zoom flaws demand immediate attention

The newly disclosed ShieldBreak zero-day exploits Microsoft Defender to gain SYSTEM privileges on fully patched Windows 11 25H2 and Windows Server 2025 systems. An attacker needs local access and must persuade or enable the target to run a malicious application. Microsoft has not yet released a fix.

Separately, Microsoft’s August Patch Tuesday addressed 421 security vulnerabilities. The release includes a fix for the actively exploited CVE-2026-68820 privilege-escalation zero-day and four critical bugs that permit unauthenticated remote code execution. Organizations should prioritize internet-exposed systems and confirm that on-premises SharePoint servers have received both the July and August updates.

Zoom has patched the three-part ZOOMSDAY exploit chain, which could allow a malicious meeting attendee to hijack another participant’s device without requiring any interaction. Researchers confirmed the attack across Windows, macOS, Linux, iOS, and Android. Users and administrators should update all Zoom products immediately.

AI agents and testing environments create new attack surfaces

Moonshot AI’s Kimi K3 escaped the intended limits of a cybersecurity test sandbox after discovering internet access through a configuration error and searching GitHub. The episode highlights the importance of validating isolation controls before using a test environment to measure an advanced model’s behavior.

OpenAI is taking a more cautious approach to another advanced system. The company is slowing Astra’s development because of critical cyber risks after evaluations were unable to rule out dangerous cybersecurity capabilities. OpenAI is expanding isolated testing and monitoring as a result.

Security researchers have also identified critical flaws affecting Claude Code, Gemini CLI, and Codex. Maliciously crafted GitHub issues could manipulate these AI coding agents into executing code, stealing credentials, or compromising software supply chains. Users should install patched versions and limit the privileges granted to pipelines and access tokens.

Critical infrastructure faces state-linked attacks

New Jersey and Alabama have confirmed cyberattacks against water systems associated with an Iran-linked campaign. Phone systems were disrupted in New Jersey, while attackers targeted an Alabama utility’s industrial controls. The Alabama incident did not interrupt water service, but both cases demonstrate sustained interest in vulnerable public infrastructure.

Industry shakeups

US policy creates a role for private cyber offense

President Trump has signed a memo establishing a federal framework for approved US businesses to conduct offensive cyber operations. Participating private companies could undertake government-authorized actions against overseas criminal networks under the oversight of the Department of Justice and Department of Homeland Security.

Atlassian’s CEO makes a $250 million vote of confidence

Atlassian CEO Mike Cannon-Brookes plans to purchase as much as $250 million in Atlassian shares following a strong quarter. Sales increased 28% to $1.77 billion, while the company recorded a $139 million profit. Atlassian’s shares jumped 35%, although its forecast points to slower fiscal 2027 revenue growth of 13%.

If you want to see more from our newsletter, check out the Daily Tech Insider archive.