Top Tech News Today, September 18, 2026: Anthropic, Google, Meta, Nvidia, OpenAI & More

It’s Friday, September 18, 2026, and the last 24 hours made one thing hard to miss: AI is being handed real authority faster than anyone is locking the doors behind it. Nvidia’s Jensen Huang said the company expects to sell twice as many chips next year. OpenAI dropped a legal-research index spanning 230 million URLs. Apple let coding agents drive Safari. Google gave households a shared family agent. Europe moved to switch AI companions off for children by default.

Then came the warning signs. A tiny startup used Claude to reach OpenAI employee accounts. A “pinned” plugin became a zero-click backdoor across four coding agents. A compromised marketing script spread malicious code across more than 100,000 websites. Spain recorded its first data breach attributed to an autonomous agent. And three CEOs persuaded Washington to abandon an industry-funded AI referee.

Meanwhile, Anthropic is putting AI into a physical biology lab, Abu Dhabi is bringing agentic AI into executive government meetings, and researchers are showing how frontier models can compress sophisticated cyber operations from months into days.

The fight is no longer simply about whether AI works. It is about who gets to act, who gets to watch, and what happens when action comes before safeguards.

That tension runs through today’s biggest technology stories. AI is moving from something people consult to something that can experiment, coordinate, predict, attack, advise, and act — while governments and companies race to build the chips, quantum systems, security controls, and rules needed to keep up.

Here are the top technology news stories shaping that transition today.

Google Launches CC, a Gemini-Powered AI Agent Built to Run Family Life

Google is testing a new AI agent called CC that can coordinate schedules, documents, email, reminders, forms, and shared household tasks for up to six family members. Unlike a standard Gemini conversation, CC gets its own Google account and can connect to information that household members explicitly choose to share. Users can route certain emails to the agent, interact with it through Google Chat, or give it access to a shared Drive folder containing invitations and documents. Each morning, CC can generate a shared “Your Day Ahead” briefing, update calendars, track unfinished tasks, create documents, and prefill forms such as school permission slips.

Google says each CC deployment runs in an isolated cloud environment using its Antigravity agent infrastructure and current Gemini models, including Gemini 3.8 Flash. Crucially, the agent can build shared memory about household preferences while maintaining limits around what information individual family members choose to expose. Google also says CC will seek confirmation before sending information outside the family group. For now, the experiment is limited to adults using personal Google accounts and is available through a waitlist. The product illustrates where consumer AI is heading: away from standalone prompts and toward persistent agents with memory, permissions, identity, and the ability to act across multiple services.

Why It Matters: Google is testing whether the next major consumer-AI interface is not another chatbot, but a persistent digital operator embedded in everyday household workflows.

Source: Ars Technica.

Microsoft and OpenAI Copyright Fight Escalates After Newly Unredacted AI Training Filings

Newly unredacted material in The New York Times’ copyright litigation against Microsoft and OpenAI has exposed internal discussions that could intensify the legal fight over AI training data. According to material described in the plaintiffs’ court brief, Microsoft executives internally worried about the economic damage AI-generated answers could cause publishers, while OpenAI personnel discussed the increasingly substitutive nature of chatbot answers. Microsoft’s own data, according to the filing, showed Copilot-generated answers cutting click-through traffic to the Times by as much as 93% compared with traditional Bing search. TechCrunch cautioned that much of the new information comes from the plaintiffs’ brief while underlying exhibits remain sealed, meaning some excerpts lack their full original context.

The filings also allege large-scale use of publisher material in training datasets, including more than 91,000 copies of works from the Times, New York Daily News, and Center for Investigative Reporting in certain mid-training datasets. Another Common Crawl-derived dataset allegedly contained more than two million documents from nytimes.com. Those assertions have not yet settled the core legal question: whether training a model on copyrighted material qualifies as fair use. But the newly disclosed material could matter because fair-use disputes often turn partly on market substitution and economic harm. Microsoft and OpenAI did not respond to TechCrunch’s requests for comment on the newly disclosed material.

Why It Matters: The AI copyright battle is increasingly shifting from abstract arguments about training data to evidence about how AI products may affect the economic market for original content.

Source: TechCrunch.

Meta Oversight Board Orders Removal of AI Deepfakes and Calls for Stronger Safeguards

Meta’s Oversight Board has ordered Facebook to remove two AI-generated videos after finding that the platform’s existing policies were insufficient in those cases. One video falsely depicted a Scottish Labor councilor making inflammatory remarks about refugees; another manipulated the image of a young Muslim campaign volunteer. The board concluded that the first violated Meta’s hateful-conduct rules and the second breached its bullying and harassment policies. It also said the material should have received stronger AI-related warnings and recommended policy changes affecting how high-risk manipulated content is labeled and distributed.

The quasi-independent board proposed nine changes, including reducing algorithmic distribution of material marked as high-risk AI content, introducing warning screens before users view some manipulated media, increasing penalties for accounts that repeatedly distribute deceptive deepfakes, and broadening Meta’s definition of unwanted manipulated imagery. The rulings illustrate a growing enforcement problem for social platforms: generative AI can create convincing fake speech or behavior at negligible cost, while content-review systems still have to determine context, intent, harm, and whether manipulation crosses an existing policy line. For platforms operating at Meta’s scale, labeling alone may prove insufficient if recommendation algorithms continue pushing misleading material to millions of users before moderation decisions are made.

Why It Matters: AI deepfakes are forcing social platforms to decide whether disclosure labels are enough or whether manipulated content also needs reduced distribution and tougher enforcement.

Source: The Guardian.

OpenAI Employee Accounts Breached in Ethical Hack Powered by Rival AI Models

Security researchers at Hacktron AI compromised several OpenAI employees’ ChatGPT accounts during an authorized security investigation, demonstrating how generative AI is lowering the barrier to sophisticated offensive-security work. The researchers initially used Anthropic’s Claude to help exploit an attack path involving an OpenAI staff discussion forum running on Discourse. From there, they reached employee accounts and progressed far enough to make a harmless pull request against OpenAI code hosted on GitHub. Hacktron said it could access code but did not download it. The team reported the work through OpenAI’s bug-bounty program, and OpenAI said it had addressed the vulnerabilities.

The striking part is not simply that an AI company was hacked during security testing. Hacktron said AI compressed work that once might have required a well-resourced team and months of effort into days. The researchers reportedly used Claude during the initial attack and later relied heavily on OpenAI’s own GPT-5.6 Sol. That creates an increasingly unusual cybersecurity loop: advanced models are simultaneously targets, defensive tools, and offensive research assistants. As AI agents gain better coding, browsing, reasoning, and tool-use capabilities, security teams may need to assume attackers can automate larger portions of reconnaissance, exploit development, and attack-chain construction.

Why It Matters: AI is shrinking the time and expertise required to execute complex cyberattacks, forcing defenders to rethink security economics.

Source: The Guardian.

Anthropic Says Claude Now Leads 26% of the Work Building Its Next AI Models

Anthropic says Claude is increasingly participating in the research required to build future versions of itself. According to the company, Claude led about 26% of its model research and development work by August, meaning it could complete most assigned tasks from a high-level prompt while remaining under human supervision. That figure was effectively zero in February. Anthropic also says roughly 90% of its research and development now involves some form of collaboration with Claude, although that broader category includes tasks performed under close human direction rather than autonomous research.

The distinction matters because claims that AI is “building itself” can easily outrun what the evidence actually shows. Anthropic says its models are not autonomously designing and training their successors. But the direction is notable: AI systems are taking on larger pieces of AI research themselves, potentially shortening the feedback loop between one model generation and the next. Anthropic disclosed that roughly 30,000 agents were performing research and engineering work as of August and has called for standardized industry metrics showing how much frontier-model development is being done by AI. The company argues that better public measurement is needed as systems move closer to forms of recursive improvement that could become harder for humans to supervise.

Why It Matters: The speed of AI development could accelerate sharply if frontier models become increasingly capable participants in the research that produces their successors.

Source: Associated Press.

Nvidia CEO Says AI Chip Demand Will Double Sales Again in 2027

Nvidia chief executive Jensen Huang told reporters on the sidelines of a King Charles III AI summit in Scotland that the company expects to sell twice as many chips next year as it does this year. Huang tied the forecast to what he described as broad, country-by-country demand, saying governments and companies in nearly every market where Nvidia operates want to invest in artificial intelligence. The comments landed a day after Nvidia shares rose more than 2 percent and as investors weighed whether the AI infrastructure cycle is cooling. Nvidia does not publish unit volumes, but Huang said last fall that the company had shipped 6 million Blackwell GPUs over four quarters. The portfolio now spans data-center Blackwell and Rubin GPUs, networking and optical silicon, CPUs, robotics and automotive processors, and the chip inside Nintendo’s Switch 2.

The forecast builds on an already aggressive financial outlook. Nvidia recently reported $96.2 billion in quarterly revenue, up 106 percent year over year, and has guided to about 70 percent growth in the fiscal year ending January 2028 — a figure that would imply roughly $673 billion in sales if supply holds. Management has separately said unconstrained customer demand points closer to a doubling, with high-bandwidth memory remaining the binding constraint through fiscal 2028. Huang used the same appearance to push back on calls for a slowdown in frontier-model development, arguing that unsafe products should not be released rather than that new laws are required. For startups building on Nvidia systems, the message is that the bottleneck is still silicon and memory, not a collapse in orders.

Why It Matters: Huang’s unit forecast keeps the AI capex debate centered on supply, not vanishing demand, and that is the variable that sets pricing power for chips, cloud capacity, and infrastructure startups.

Source: CNBC.

Zuckerberg, Musk and Huang Press Trump to Kill Industry-Funded AI Regulator

Mark Zuckerberg, Elon Musk and Jensen Huang separately urged President Donald Trump in recent weeks to reject an industry-funded AI oversight body modeled on ideas floated by Google DeepMind chief Demis Hassabis, according to reporting that circulated Thursday. The proposed organization would have resembled a FINRA-style standards-and-supervision outfit for frontier models. The three executives argued that such a body would concentrate power in the hands of OpenAI, Anthropic and Google and raised questions about who would sit on it. White House officials have told other AI executives that building consensus inside the administration is difficult because opponents of tighter rules are calling Trump directly.

The lobbying landed in the same week that Anthropic’s Dario Amodei called for a slower pace of capability gains, with Sam Altman and Hassabis publicly aligning on the need for more caution while Huang and Zuckerberg opposed a pause. Trump has dismissed existential-risk warnings as a “hoax” and argued that heavy U.S. restrictions would give China an advantage. Huang said at Salesforce’s Dreamforce conference that existing product-liability and safety laws are enough: if a company is not confident in a system, it should not ship it. Musk has instead floated mutual audits in which rival labs test one another’s models. Startups that expected a single private-sector rulebook now face a split market: safety-focused labs pushing standards, and the largest platform and chip companies defending a lighter federal touch.

Why It Matters: The collapse of a shared industry regulator leaves AI policy dependent on personal access to the White House, raising the odds of uneven rules for labs, chipmakers, and the startups that build on both.

Source: TechStartups via The Wall Street Journal, Forbes

Anthropic Quietly Builds Physical Biology Lab as AI Push Moves Into Drug Science

Anthropic has quietly established a wet lab in the San Francisco Bay Area, pushing its AI ambitions beyond computer simulations and into physical biological experiments. The company’s head of life sciences, Eric Kauderer-Abrams, confirmed the facility to Reuters, saying real-world laboratory work remains essential for testing biological ideas. Anthropic is also exploring how its Claude models could direct robotic lab equipment with limited human intervention, although the company says human oversight remains central. Anthropic clarified that the facility itself is not specifically a drug-discovery lab, but life sciences has become one of the company’s largest investment areas by headcount and resources.

The move is significant because frontier AI companies are beginning to cross the boundary between generating scientific hypotheses and physically testing them. Anthropic has already launched Claude Science, acquired Coefficient Bio, and works with pharmaceutical companies including Genentech, Bristol Myers Squibb, and Novo Nordisk. Its longer-term focus includes difficult or commercially neglected diseases, while the company says it does not plan to run clinical trials and does not want to compete directly with drugmakers. The larger question is whether AI labs can compress the experimental cycle itself by connecting models, robotics, laboratory instrumentation, and biological data into one system.

Why It Matters: AI is beginning to move from analyzing scientific data to participating directly in the physical process of experimentation.

Source: Reuters.

AI Forecasting Startup Mantic Raises More Than $20M After Beating Human Forecasters

London-based ManticAI has raised more than $20 million after its forecasting system outperformed human participants in a major prediction competition. Sky News reported that Balderton Capital and Radical Ventures jointly led the financing, with Microsoft’s M12 venture fund also participating. The startup was co-founded by former Google DeepMind researcher Toby Shevlane and Ben Day and specializes in adapting frontier AI models to assign probabilities to future political, economic, business, and cultural events. Sky reported that the round was being raised at a valuation close to $90 million.

Mantic drew attention after its system performed strongly in the 2026 Metaculus forecasting competition, where AI systems collectively showed they could compete with — and in some cases outperform — experienced human forecasters. Forecasting represents an interesting next step for enterprise AI because the output is measurable: a system assigns probabilities, the event happens or does not, and performance can be scored over time. That makes forecasting potentially useful in markets, supply chains, risk management, strategic planning, and intelligence analysis. It also creates a tougher standard than fluent text generation because impressive language cannot compensate indefinitely for poor calibration. Commercial success, however, will depend on whether competition performance carries over reliably to messy real-world decisions.

Why It Matters: If AI forecasting continues improving, probabilistic prediction could become a new enterprise-AI category alongside coding, research, and data analysis.

Source: Sky News.

France Opens Smart-Glasses Probe as AI Wearables Trigger New Privacy Backlash

French prosecutors have opened a criminal investigation into suspected sexual harassment involving smart glasses, while France’s data-protection authority is fielding complaints and questions from businesses about the devices. Prosecutors told Reuters that at least one investigation followed complaints involving a social-media trend in which women were secretly filmed on the street using smart glasses and the resulting footage was posted online. Authorities did not identify the brand involved. France’s CNIL privacy regulator separately said it had received fewer than 10 complaints involving smart glasses in workplaces and was increasingly being asked by employers whether they could prohibit the devices.

The scrutiny arrives as camera-equipped AI glasses move from niche hardware into the consumer mainstream. Meta’s glasses made with EssilorLuxottica hold roughly three-quarters of the global market, according to Reuters, and sold about seven million units last year. But discreet cameras, microphones, AI assistants, and always-available recording create social questions that smartphones largely avoid because a phone is usually visible when someone points it at another person. Australia is also considering restrictions on camera-equipped smart glasses in government workplaces, while companies and institutions elsewhere are reviewing their own rules. The industry’s challenge is becoming clear: the better smart glasses become at blending into ordinary eyewear, the harder it becomes for bystanders to know when sensors are active.

Why It Matters: Smart glasses may be approaching a regulatory inflection point as wearable AI collides with consent, workplace security, and privacy expectations.

Source: Reuters.

India Deep Tech Alliance Deploys ₹2,170 Crore Across 56 AI and Frontier-Tech Startups

Members of the India Deep Tech Alliance have invested nearly ₹2,170 crore across 56 deep-tech companies during the alliance’s first year, providing a snapshot of how private capital is moving into India’s strategic technology sectors. AI and AI applications attracted about ₹639 crore across 16 companies, while quantum computing, robotics, and space drew roughly ₹649 crore across 21 companies. Another ₹655 crore went to energy security, climate, and energy-transition companies, while biotechnology and related sectors received about ₹189 crore. Alliance members made the investments independently between September 2025 and August 2026.

The alliance includes investors such as Accel, Blume Ventures, 3one4 Capital, Celesta Capital, and Chiratae Ventures, whose members have collectively committed more than $2.5 billion over five years to AI and frontier technology. At SEMICON India, the group also announced plans with industry association SEMI for an India Semiconductor Academy intended to turn the country’s large engineering base into more job-ready chip talent. That effort matters because semiconductor ambitions cannot be met through fabrication subsidies alone; design engineers, equipment expertise, packaging skills, process knowledge, and specialized technical training are equally important. India is increasingly treating deep tech as infrastructure rather than simply another venture category.

Why It Matters: India’s deep-tech push is moving beyond government incentives into sustained private investment across AI, chips, quantum, robotics, space, biotech, and energy.

Source: Business Standard.

L&T Semiconductor Unveils 40-Chip Portfolio as India Pushes for Tech Sovereignty

L&T Semiconductor Technologies has unveiled a 40-product portfolio at SEMICON India 2026 spanning compute, connectivity, security, power electronics, surveillance, and AI acceleration. The Larsen & Toubro subsidiary highlighted two technical milestones: its first 1,200-volt silicon-carbide MOSFET platform and the tape-out of a BLDC motor-controller chip fully designed in India. Its compute portfolio also includes a vision-camera system-on-chip with on-device AI acceleration, while other products target secure identity, smart metering, automotive systems, digital payments, data centers, and critical infrastructure.

The announcement reflects a broader shift in India’s semiconductor strategy. Building fabs remains important, but domestic chip sovereignty also requires local intellectual property and products deployable in real systems. Silicon-carbide devices are increasingly valuable in EV chargers, microgrids, traction inverters, and other high-efficiency power systems, while locally designed controllers and AI-capable edge chips could reduce reliance on imported components in high-volume applications. The challenge will be moving from prototypes and tape-outs to competitive manufacturing, yield, packaging, qualification, and commercial scale. Still, the breadth of L&T’s portfolio signals that India’s semiconductor effort is widening from assembly and manufacturing incentives into indigenous product design.

Why It Matters: India’s chip strategy is increasingly focused on owning more of the semiconductor stack — from design and intellectual property to manufacturing and end products.

Source: TimesTech.

U.S., Japan, Australia and Germany Attribute Global Tech-Recruiting Attacks to North Korea-Linked WaterPlum

Japanese authorities, working with agencies in the United States, Australia, and Germany, have publicly attributed a large cyber campaign targeting technology professionals to the North Korea-linked group WaterPlum, also widely known as Contagious Interview. Japan’s National Police Agency said the attackers approached job seekers and software developers through apparently legitimate recruiting or technical-interview processes before attempting to compromise their computers. Authorities said the operation targeted people in Japan, the United States, Europe, and other regions and was used to steal sensitive information and cryptocurrency.

Authorities described the scale as substantial. MLex reported that the joint assessment linked the operation to infections of at least 30,000 devices across more than 100 countries and to roughly $10.7 million in cryptocurrency transferred to North Korea. Japanese reporting based on the advisory said the attackers stole more than 7,000 cryptocurrency-wallet records. The campaign exploits an unusually effective trust mechanism: developers expect recruiters and prospective employers to ask them to download repositories, test software, or run code during technical interviews. That creates an attack surface conventional email-security systems cannot easily eliminate. The campaign also overlaps with broader warnings about North Korean IT workers seeking legitimate remote employment abroad to generate foreign currency.

Why It Matters: Software hiring itself has become a supply-chain security risk as state-linked attackers exploit technical interviews to get malicious code onto developers’ machines.

Source: Japan National Police Agency and MLex.

Brevo Supply-Chain Attack Injects Malicious Code Into More Than 100,000 Websites

Customer-engagement platform Brevo disclosed a supply-chain attack that ultimately delivered malicious JavaScript through more than 100,000 websites. The incident began September 10, when an attacker exploited a flaw involving Brevo’s handling of SAML single sign-on and accessed 138 customer accounts. Brevo said phishing emails were sent from six accounts and contacts were exported from 43. After the initial intrusion was contained, the attackers returned on September 14 using a compromised long-lived Cloudflare API key to deploy a malicious Cloudflare Worker.

That Worker injected malicious code into Brevo-owned domains and into JavaScript files that customers embed on their own websites. Some visitors were shown a fake Cloudflare verification page instructing them to copy and run a command on their computers — a social-engineering technique commonly called ClickFix. On WordPress sites using Brevo widgets, the script could also attempt to deploy a plugin when a logged-in administrator visited the compromised page. The episode demonstrates why third-party web components are increasingly attractive targets: compromising one trusted service can distribute malicious code across thousands of otherwise unrelated organizations without attackers having to breach each site individually. Long-lived cloud credentials also remain especially dangerous because they can survive after the original intrusion path closes.

Why It Matters: Modern web infrastructure creates enormous attack multipliers — a single compromised SaaS provider can expose users across more than 100,000 downstream sites.

Source: SecurityWeek.

U.S. Launches $215M Quantum Competition Targeting Fault-Tolerant 100-Qubit Systems

The U.S. Department of Energy has launched the Quantum Genesis Q Competition, offering up to $215 million in planned funding to private companies that can demonstrate scientifically relevant fault-tolerant quantum computers. The program targets machines with at least 100 logical qubits that can perform hundreds of millions of fault-tolerant operations. Phase one offers milestone awards of up to $1.5 million per participant. Phase two includes a $100 million incentive pool for systems reaching the 100-logical-qubit threshold, plus separate $50 million bonus pools for demonstrations reaching 150 and 200 logical qubits.

DOE is separately planning a $45 million national-laboratory validation and verification program to test the systems across hardware, logical architectures, algorithms, controls, and applications. That testing component matters because quantum-computing announcements often rely on metrics that are hard to compare across architectures. Logical qubits — error-corrected units capable of carrying useful computation — offer a more demanding milestone than raw physical-qubit counts. DOE says the target systems should eventually address scientific problems in chemistry, materials, physics, and applied mathematics that remain difficult for classical machines. Funding beyond the initial fiscal-year allocation remains subject to congressional appropriations, so the full $215 million is planned, not guaranteed.

Why It Matters: The quantum race is shifting from demonstrations of physical qubit counts toward error-corrected machines that can perform verifiable scientific work.

Source: The Quantum Insider.

China’s Leading AI Models Generate About One-Tenth the Revenue of OpenAI and Anthropic

China’s leading AI developers are building globally competitive models, but new estimates highlight a large commercialization gap with their biggest U.S. rivals. Seven Chinese developers — ByteDance, Alibaba, DeepSeek, Moonshot AI, Z.ai, MiniMax, and Kuaishou — generated an estimated combined $10.7 billion in annual recurring revenue measured across recent months, according to Rhodium Group figures reported by the South China Morning Post. That compares with more than $100 billion in combined reported ARR for OpenAI and Anthropic. ByteDance led the Chinese group with roughly $4 billion in ARR, followed by Alibaba at about $2.4 billion.

ARR is a run-rate metric rather than audited full-year revenue, so treat the comparison as an indicator of commercial scale rather than a direct accounting comparison. Even so, the gap illustrates an important distinction in the global AI race: benchmark performance and model capability do not automatically translate into monetization. U.S. frontier labs have benefited from large enterprise contracts, expensive premium subscriptions, API demand, and deep cloud-distribution partnerships. Chinese developers often operate in a more price-competitive domestic market and have increasingly emphasized open-weight models, which can expand adoption without generating equivalent direct revenue. The result could be two different AI business models emerging at scale: one optimized for monetization and another for distribution and ecosystem reach.

Why It Matters: The next phase of the global AI race will be measured as much by sustainable revenue and distribution as by model benchmarks.

Source: South China Morning Post.

Abu Dhabi Uses Agentic AI Inside Executive Council Meeting for the First Time

Abu Dhabi’s Executive Council has used an agentic AI platform during a council meeting for the first time, giving AI a direct supporting role in high-level government deliberations. The AI Committee System analyzes proposals and agenda items before meetings, produces executive summaries, and can generate real-time questions, fact checks, observations, and analytical recommendations while officials discuss policy. Gulf News reported that outputs remain traceable to their original source documents, an important feature when AI-generated analysis is being introduced into government decision workflows.

The system also tracks meeting proceedings, creates analytical reports and minutes, summarizes recommendations, and assigns follow-up responsibilities. During the latest meeting, officials considered programs covering infrastructure, childcare, reproductive health, and AI education, among other areas. The deployment is notable because most government AI initiatives to date have concentrated on public-facing chatbots, administrative automation, document processing, or citizen services. Abu Dhabi is moving the technology further upstream into the machinery that prepares and supports executive decisions. That also raises the stakes for provenance, auditability, cybersecurity, model reliability, and human accountability. Agentic systems that can produce recommendations or check claims can improve information processing, but officials still need clear visibility into the evidence and assumptions behind every output.

Why It Matters: Agentic AI is moving from automating government paperwork to supporting the decision-making process itself.

Source: Gulf News.