Top Tech News Today, September 2, 2026: Anthropic, Google, Meta, Nvidia, Perplexity, OpenAI, TenCent & More

It’s Wednesday, September 2, 2026, and today’s tech news is less about new toys than new limits. OpenAI just rated its next model a critical cyber risk. Texas froze data-center power hookups after a rush of “ghost” demand. New York City barred student AI through eighth grade. Uber cut 3,300 jobs to fund robotaxis. In the same window, Anthropic made AI agents cheaper, Apple’s new CEO teased a “huge launch,” and India moved to let software spend money on UPI.

The common thread is hard to miss: AI is getting cheaper, more capable, and more deeply embedded in the systems that move money, consume electricity, secure networks, shape classrooms, and increasingly make decisions without humans in the loop.

From Anthropic and Astra to Texas, UPI, Apple, and Uber, here are the top tech stories making waves today.

Uber lays off 3,300 Staff, 10% of Global Workforce, as It Flattens Management and Redirects Resources

Uber is cutting about 3,300 jobs, roughly 10% of its global workforce, as CEO Dara Khosrowshahi moves to reduce management layers, consolidate teams and simplify an organization that expanded significantly over the past several years. The company said it has notified affected workers, except where local employment rules require a different process. Uber is also concentrating more employees in major hubs and reducing fully remote positions.

The restructuring is noteworthy because Uber says the cuts are not the result of a collapsing core business. Instead, management argues that organizational complexity has slowed decision-making and produced too many small teams and layers of coordination. The company is simultaneously facing an expensive technological transition as autonomous vehicles move closer to commercial deployment. Waymo and other robotaxi developers could eventually alter ride-hailing economics, forcing Uber to decide how much technology it builds itself and how much it accesses through partnerships. Cutting bureaucracy while redirecting investment toward autonomous transportation makes the layoffs part of a broader strategic repositioning, not a conventional downturn response.

Why It Matters: Uber’s restructuring shows how mature tech companies are cutting organizational complexity even as they invest heavily in AI and autonomous technologies that could reshape their existing businesses.

Source: TechStartups via Uber, Reuters.

Nvidia Backs Spanish Startup iPronics in $125 Million AI Networking Round

Spanish semiconductor startup iPronics has raised $125 million from investors including Nvidia to develop networking technology for AI data centers. The Valencia-based company specializes in programmable photonic chips that use light rather than conventional electrical signaling to move data between computing systems. As AI clusters grow from thousands to potentially hundreds of thousands of accelerators, networking is becoming one of the hardest engineering constraints in large-scale AI infrastructure.

That makes companies such as iPronics strategically interesting to Nvidia even though they operate outside the GPU itself. Faster accelerators deliver diminishing returns when processors cannot exchange information quickly enough, turning interconnects, switches, optics and memory into critical pieces of the AI-computing stack. Photonics could eventually reduce power consumption and increase bandwidth compared with traditional electrical connections, although moving new optical architectures into commercial data centers remains technically difficult. The investment also shows AI infrastructure capital spreading beyond headline chipmakers into the less visible components needed to build enormous clusters. Europe, meanwhile, gains a potentially important semiconductor company in a segment increasingly tied to AI sovereignty.

Why It Matters: The AI infrastructure race is shifting from simply acquiring GPUs to solving the networking bottlenecks that determine how efficiently massive clusters actually operate.

Source: Reuters.

Anthropic Launches Claude Fable 5.1 and Mythos 5.1 With 45% Lower AI Agent Costs

Anthropic has released Claude Fable 5.1 and Claude Mythos 5.1, its latest generation of models for coding, knowledge work, cybersecurity, and scientific research. Fable 5.1 is broadly available, while the more capable Mythos 5.1 remains restricted to vetted organizations because of its advanced cybersecurity and biology capabilities. Anthropic says the new Fable can deliver similar or better performance than its predecessor at lower reasoning settings while cutting the cost of some highly agentic workloads by up to 45%.

Cost could prove as significant as benchmark improvements. Enterprises increasingly care about the cost of sustained agentic work, not how impressively a model answers isolated prompts. Anthropic has reduced cache-read pricing and tightened safeguards, making it less likely that legitimate cybersecurity and scientific requests will be blocked unnecessarily. Fable 5.1 can now help identify software vulnerabilities in source code, although Anthropic continues to reserve more sensitive exploit-generation capabilities for restricted models. That increasingly visible separation between broadly available AI and controlled frontier capability may become a template for the industry as model developers confront systems that are economically valuable but potentially dangerous when deployed without restrictions.

Why It Matters: The next stage of AI competition is increasingly about the combination of intelligence, agent economics, privacy, and safety rather than benchmark scores alone.

Source: The Verge.

Google Brings Gemini Memory to Android Find Hub in September Feature Drop

Google’s September Android update is adding a Gemini-powered capability that lets users tell their phones where physical objects have been stored. Instead of attaching a Bluetooth tracker to everything, users can say that a passport, key or other item was placed in a particular location and optionally save a photo. Gemini stores that information inside Find Hub so it can be retrieved later through voice or the app’s remembered-items section. Google is also rolling out Motion Assist for reducing motion sickness and camera-based Guided Vision features.

The update offers a small but telling example of how generative AI is becoming embedded into ordinary operating-system functions rather than remaining inside standalone chatbot interfaces. Remembering where an object was placed is not computationally spectacular, but it turns AI into persistent contextual infrastructure that can connect conversations with real-world actions. As Google integrates Gemini across Android, Search, Workspace and other products, the competitive advantage increasingly comes from distribution and context rather than the model alone. Apple, Microsoft and Meta are pursuing similar strategies across their own device ecosystems.

Why It Matters: Consumer AI is beginning to disappear into everyday operating-system features, where persistent context may prove more useful than another standalone chatbot.

Source: The Verge.

Researchers Build a $7 Phone Clip-On That Spots Hidden Cameras With 94% Accuracy

A research team unveiled a smartphone clip-on that costs about $7 to build and uses an AI model plus a dynamic LED grid to find hidden cameras, reporting 94% accuracy in tests. The accessory mounts on a phone camera and pulses structured light while onboard software looks for the optical signature of a lens rather than relying on a user to sweep a room by eye. That combination is meant to catch devices tucked in smoke detectors, clock radios, USB chargers, and other household objects that have become standard props in short-term rental and workplace privacy scares.

Consumer “camera finders” already sell on marketplaces, but most are either expensive infrared gadgets or apps that merely hunt for open Wi-Fi streams. A cheap, repeatable optical method that runs on a phone is a different product category: it could ship as a travel accessory, an enterprise inspection tool, or a feature bundled into future phone hardware. False positives and lighting conditions will decide whether the 94% lab figure survives hotel rooms and offices. Even so, the prototype lands in a year when cheap cameras and AI-altered video have made physical-space verification part of consumer security, not only national-security kits.

Why It Matters: A few dollars of optics and on-device AI could turn hidden-camera detection from a specialist gadget into a default travel and workplace tool.

Source: Tom’s Hardware.

Tencent-Backed AI Chipmaker Enflame Draws Massive Demand for Shanghai IPO

Chinese AI chipmaker Enflame Technology drew extraordinary investor demand for its Shanghai listing, with its online offering reportedly oversubscribed more than 6,000 times. The Tencent-backed company is raising roughly 6.1 billion yuan, or about $900 million, as China continues pouring capital into domestic alternatives to Nvidia and other foreign semiconductor suppliers. Enflame develops GPUs and accelerators aimed at AI training and inference workloads, putting it directly inside one of Beijing’s most strategically important technology sectors.

The demand illustrates how semiconductor independence has become both an industrial-policy objective and an investment thesis in China. U.S. export controls have made access to the most advanced American AI chips more difficult, creating incentives for Chinese cloud providers, model developers, and government-backed buyers to adopt locally produced hardware. Enflame still faces formidable technical and ecosystem competition from Nvidia, Huawei, and other domestic chip developers, and strong IPO demand does not establish technological parity. But access to public capital can give chip startups the billions needed for research, software development, and manufacturing commitments. China increasingly appears determined to build an entire parallel AI-compute stack rather than remain dependent on imported accelerators.

Why It Matters: Enflame’s IPO demand shows investors are treating China’s push for domestic AI computing infrastructure as a long-term strategic market, not a temporary response to export restrictions.

Source: Reuters.

OpenAI Says Astra Has Crossed a Critical AI Cybersecurity Threshold

OpenAI says its forthcoming Astra model has reached a level of cybersecurity capability that requires stronger safeguards before broader deployment. Under OpenAI’s Preparedness Framework, Astra now qualifies as having “Critical” cyber capability, meaning the model can potentially identify previously unknown vulnerabilities and develop ways to exploit them across well-defended systems with substantially less human guidance than existing public models. The company said the assessment triggered additional controls around access, monitoring, model behavior, and deployment.

The disclosure offers an unusually concrete glimpse at where frontier AI development is heading. Coding assistants are already common tools for software development and defensive security, but models that can autonomously discover and exploit vulnerabilities change the risk equation. The same systems that can help defenders find weaknesses faster could dramatically lower the expertise and labor needed to conduct sophisticated attacks. OpenAI plans to restrict Astra initially rather than make all its capabilities broadly available. The development also increases pressure on other frontier labs to explain how they will handle models whose offensive cybersecurity abilities begin approaching those of skilled human teams.

Why It Matters: Astra shows the AI race is moving beyond better chatbots and coding assistants into systems that can independently perform security work that previously required highly specialized human expertise.

Source: The Information / OpenAI.

AI Security Startup AIR Raises $50 Million to Police the Agent Software Supply Chain

AI security startup AIR has emerged from stealth with $50 million in funding to secure the growing ecosystem of tools used by autonomous AI agents. The company, founded by former Israeli Unit 8200 cybersecurity specialists Yair Saban and Niv Hoffman, raised the capital across two seed rounds led by Sequoia and Greenoaks. AIR says its platform can discover AI agents operating inside enterprises, inspect the plug-ins, skills and MCP servers they use, and block components that fail security checks.

The problem AIR is targeting could become an important new layer of enterprise security. AI agents increasingly connect to email, databases, cloud systems, browsers and internal applications, often through third-party extensions. That creates an emerging software supply chain where a malicious or compromised component could manipulate an agent into exposing information or executing damaging actions. AIR says roughly 27% of the publicly available add-ons and skills it evaluates fail its security criteria. Competition is already forming around the category, including companies such as Noma Security, Zenity and Astrix. The funding suggests investors expect agent governance to become a security market of its own rather than simply another feature inside existing endpoint products.

Why It Matters: As AI agents gain permission to act inside corporate systems, securing what those agents can install, access and trust may become as important as securing human users.

Source: TechCrunch.

Perplexity Launches Hybrid AI That Keeps Sensitive Work on the Mac

Perplexity has launched a hybrid-compute capability for its Computer agent that can divide a single AI task between cloud models and locally running models on Apple silicon Macs. Tasks can begin using more capable frontier models in the cloud, while steps involving confidential files or sensitive information can be handed to an open-weight model running directly on the user’s machine. Perplexity says the transition happens without restarting the workflow or losing the agent’s context.

That architecture addresses one of the biggest objections enterprises have to autonomous AI: data exposure. Companies may want powerful cloud models for difficult reasoning while refusing to send contracts, financial records, source code or internal documents outside their own hardware. Hybrid inference offers a potential compromise, especially as personal computers become powerful enough to run increasingly capable models locally. It also creates a new competitive dimension for Apple silicon and other AI-oriented PC hardware. Instead of choosing exclusively between cloud AI and on-device AI, applications may increasingly orchestrate both. If that pattern spreads, model routing and privacy-aware computing could become core parts of enterprise agent infrastructure.

Why It Matters: Hybrid AI could let companies adopt capable autonomous agents without forcing every sensitive document or workflow into an external cloud.

Source: VentureBeat.

Meta Disables Cameras on Thousands of Tampered AI Smart Glasses

Meta has disabled camera functionality on thousands of pairs of its smart glasses after detecting that users had tampered with the recording indicator light. Meta told Semafor that fewer than one-tenth of 1% of glasses sold had been altered, but the company has strengthened software designed to determine whether the LED signaling that a camera is recording has been covered or physically damaged. Devices modified by drilling into the indicator can lose camera functionality, while users who simply covered the light can restore it by removing the obstruction.

The enforcement highlights an issue that will become harder as wearable AI devices become more common. Smart glasses can place cameras, microphones and AI assistants into ordinary-looking eyewear, giving users useful hands-free capabilities while making it harder for people nearby to know when recording is taking place. Meta’s decision effectively enforces a physical privacy feature through software, setting a precedent other wearable manufacturers may follow. With Apple, Google, Samsung and startups pursuing AI-powered glasses and ambient computing devices, the industry will increasingly need technical mechanisms that protect people who never agreed to interact with the device.

Why It Matters: The smart-glasses market will depend as much on establishing social trust around cameras and sensors as it does on making wearable AI technically useful.

Source: Semafor.

Aesto Health Data Breach Exposes Information on More Than 9.5 Million People

Healthcare technology provider Aesto Health says a cyberattack affected more than 9.5 million individuals, making it one of the larger healthcare-data disclosures reported this year. Aesto provides software that medical organizations use to migrate, archive, and access patient information, particularly during electronic health record transitions and healthcare acquisitions. The company said an unauthorized actor accessed part of its Amazon Web Services environment in December 2025, although investigators did not confirm the scope of potentially compromised protected health information until May 2026.

The delayed disclosure underscores why healthcare infrastructure remains an attractive target. Vendors sitting between hospitals, doctors and electronic health-record systems can aggregate enormous volumes of patient information even though most consumers have never heard of them. That creates concentrated risk across the healthcare software supply chain. Unlike passwords, sensitive medical histories and identity information cannot simply be changed after a breach. The incident also illustrates how long forensic investigations can take when attackers gain access to large cloud repositories. Organizations increasingly have to evaluate the security of the software providers storing historical patient records, not merely the systems used during active treatment.

Why It Matters: Healthcare breaches increasingly originate in the technology vendors connecting medical institutions, turning relatively obscure infrastructure companies into repositories for millions of highly sensitive records.

Source: BleepingComputer.

Hackers Are Exploiting a Critical Langflow Flaw to Steal OpenAI and AWS Keys

Attackers are actively exploiting a critical vulnerability in Langflow, the open-source framework used to build AI applications and agent workflows. The flaw, tracked as CVE-2026-0768, can allow an unauthenticated attacker to remotely execute Python code on vulnerable servers. Security researchers have observed attacks attempting to extract credentials and secrets, including OpenAI API keys, AWS credentials, and other tokens stored inside exposed Langflow environments.

The incident shows how AI development frameworks are becoming high-value cybersecurity targets. Tools such as Langflow often sit directly between models, databases, APIs, and cloud infrastructure, meaning compromised installations can give attackers something more valuable than the application itself: credentials that unlock other services. The vulnerability is especially concerning because experimentation around AI agents has encouraged developers to connect numerous services quickly, sometimes without mature production security practices. As AI applications move into enterprise deployment, development frameworks will need the same patching discipline traditionally applied to internet-facing servers, identity infrastructure, and software supply chains.

Why It Matters: AI tooling is becoming part of the attack surface, and compromising an orchestration framework can expose the cloud and model credentials connected to everything behind it.

Source: BleepingComputer.

China-Linked ‘Fire Ant’ Hackers Turn Cisco Routers Into Launchpads for Deeper Attacks

A China-based hacking operation tracked as “Fire Ant” compromised Cisco networking equipment and used the routers as footholds for spying, credential theft and attacks against additional connected environments, according to cybersecurity firm Sygnia. The campaign targeted infrastructure that organizations often regard as trusted plumbing rather than actively monitored endpoints, giving attackers a strategically valuable position inside networks.

Compromised routers are especially dangerous because they sit directly in the path of network traffic and can provide attackers visibility into authentication and communications systems. Network appliances also tend to receive less behavioral monitoring than laptops and servers, allowing sophisticated intruders to remain hidden longer. The campaign reinforces a broader shift in state-linked cyber operations toward edge devices, VPN appliances, firewalls and other infrastructure that can provide persistent access without relying on conventional malware installed on employee machines. For multinational companies and governments, securing old networking hardware is increasingly a national-security issue rather than routine IT maintenance.

Why It Matters: Attackers who control the network infrastructure itself can bypass many defenses built to protect individual computers, making routers and edge devices increasingly valuable targets for sophisticated cyber operations.

Source: The Record.

Palo Alto Networks Buys AI Agent Startup Console to Automate Security Operations

Palo Alto Networks has acquired Console, an AI-native platform built to automate enterprise workflows using agents. The cybersecurity company plans to integrate Console’s technology into its Cortex platform, where AI agents could help analyze alerts, prioritize incidents, and take actions across customer environments. Palo Alto Networks did not disclose the acquisition’s financial terms.

The purchase illustrates how the cybersecurity market is moving from AI-assisted analysis toward AI-operated security workflows. Security teams have spent years managing growing queues of alerts generated by dozens of tools. Agentic systems promise to investigate context, connect information across systems, and perform routine remediation without requiring an analyst to manually handle each step. That opportunity also comes with obvious risks: an autonomous security agent with broad privileges could cause substantial damage if compromised or misdirected. Palo Alto’s move puts agent architecture closer to the center of its security strategy and could pressure rivals to accelerate similar capabilities.

Why It Matters: Cybersecurity platforms are beginning to treat autonomous AI agents as operators inside the security stack rather than simply copilots that summarize alerts for humans.

Source: SecurityWeek.

German Space Startup HyImpulse Raises More Than €50 Million for European Rocket Launches

German space startup HyImpulse Technologies has raised more than €50 million in a Series A extension to develop its commercial launch business. The round was co-led by JOIN Capital and Ace Capital Partners, with participation from investors including North Ventures, BW-Capital, Bayern Kapital and the German Aerospace Centre. HyImpulse is developing rockets for suborbital and orbital missions and says its combined equity and public funding now exceeds €125 million.

The financing arrives as Europe tries to rebuild sovereign launch capacity and reduce dependence on non-European providers. SpaceX’s scale and launch cadence have transformed the economics of reaching orbit, leaving European companies and governments searching for competitive domestic alternatives. HyImpulse uses a hybrid propulsion approach intended to simplify manufacturing and reduce some of the costs associated with conventional liquid-fueled rockets. The company still has to prove that its systems can deliver reliable commercial launches, but continued private investment indicates growing confidence that Europe can support multiple launch providers rather than relying predominantly on government-led programs.

Why It Matters: Europe increasingly sees independent access to orbit as strategic infrastructure, creating an opening for private rocket startups that can supplement traditional government-backed launch programs.

Source: EU-Startups.

Spanish Biotech Startup iPremom Raises €15 Million for Early Pregnancy Risk Detection

Spanish biotech startup iPremom has raised €15 million in seed funding to advance a blood test intended to identify risks of pregnancy complications during the first trimester. Amadeus Capital Partners led the round, with Asabys Partners and APEX Ventures participating. The Valencia-based company’s MaiRa platform analyzes cell-free RNA signals in maternal blood to estimate the risk of complications including preeclampsia, preterm birth, fetal-growth restriction and gestational diabetes before symptoms become apparent.

The funding will support clinical validation, regulatory work and commercial deployment. iPremom says its technology is based on research involving more than 26,000 samples connected to roughly 10,000 pregnancies. Earlier detection could allow physicians to monitor high-risk patients before complications become clinically obvious, although large clinical studies and regulatory review will ultimately determine how widely such testing can be used. The company represents a broader shift in biotech toward using molecular data and computational analysis to identify disease risk earlier rather than waiting for conventional symptoms or diagnostic markers.

Why It Matters: Molecular diagnostics are moving preventative medicine further upstream, where identifying risk before symptoms develop could change both treatment decisions and healthcare economics.

Source: Sifted.