Top Tech News Today, September 23, 2026: Anthropic, AT&T, Google, Microsoft, Nvidia, OpenAI, Xiaomi & More
Nickie Louise
Posted On September 23, 2026
It’s Wednesday, September 23, 2026, and the AI boom just showed up where it usually doesn’t: on an FBI jobs portal, in a UN speech rejecting a global rulebook, and in a $500-a-month chatbot that was already emptying corporate inboxes. While AI labs cut the cost of frontier coding models and Qualcomm put 2-nanometer agentic chips into smartphones, the real battle shifted to cooling contracts, training-data factories, agent permissions, and a U.S.–China AI “hotline” that, for now, is still just two officials and a phone.
AI is getting cheaper, cyberattacks are getting smarter, and the infrastructure race behind both is becoming harder to ignore. Over the past 24 hours, Anthropic rolled out a more efficient Claude model, Microsoft helped dismantle an AI-powered phishing operation, Nvidia pushed deeper into robotics, Xiaomi unveiled a trillion-parameter model, and U.S. regulators opened a new investigation into hands-off driving technology. Meanwhile, critical security flaws, quantum tools, biotech funding, and fresh startup capital are offering a clearer picture of where technology — and the money behind it — is moving next.
From AI and startups to regulation, cybersecurity, chips, and Big Tech, here are the technology stories making waves today.
Anthropic launches Claude Opus 5.5 with lower costs and tighter AI safety controls
Anthropic has released Claude Opus 5.5, the first model in its new Claude 5.5 family, promising performance comparable to its higher-end Claude Fable 5.1 while substantially cutting operating costs. The company says Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, versus $5 and $25 for the previous Opus 5. Anthropic says the new model can handle long-running coding and knowledge-work tasks, ships with a 1-million-token context window, and includes always-on adaptive reasoning.
The timing matters because frontier-model economics are becoming almost as important as raw intelligence. OpenAI, Anthropic, Google, and Chinese AI developers are being pushed to deliver more capability per dollar as enterprises move from experiments into high-volume deployments. Anthropic also subjected Opus 5.5 to external evaluations by METR and Frontier Design and says its internal testing found fewer attempts to circumvent safety boundaries. Those findings are company-reported, but the inclusion of outside evaluators reflects growing scrutiny around increasingly autonomous models.
Why It Matters: The frontier AI contest is shifting from pure benchmark leadership toward a three-way fight over capability, operating cost, and safety.
Source: Reuters.
Nvidia launches Isaac ROS 5.0 to bring AI agents deeper into robotics development
Nvidia has released Isaac ROS 5.0, the latest version of its GPU-accelerated software stack for developers building robots with the Robot Operating System. Announced at ROSCon in Toronto, the release adds AI-agent skills intended to let coding assistants help developers configure environments, migrate software, and work with Nvidia’s robotics tools. Isaac ROS 5.0 also adds support for ROS 2 Lyrical and expands deployment support across Nvidia’s Jetson edge-computing platform.
The release highlights an important shift in physical AI: AI models are beginning to participate in building the robots that will eventually run AI models themselves. Robotics software remains fragmented and technically demanding, with developers integrating sensors, perception systems, motion planning, control software, and hardware-specific libraries. Giving coding agents standardized robotics skills could shorten development cycles and make increasingly complicated robotics stacks accessible to smaller engineering teams. Nvidia benefits regardless of which robot manufacturers ultimately dominate because Isaac, Jetson, and its broader computing stack sit underneath many robotics projects.
Why It Matters: Nvidia is extending its AI platform strategy beyond data centers and into the development layer for autonomous machines.
Source: NVIDIA Blog.
FBI investigates ShinyHunters claim of massive breach exposing employee and applicant data
The FBI is investigating claims from hacking group ShinyHunters that it stole more than 2 terabytes of sensitive data connected to FBI employees and job applicants. The group claims the information includes names, contact details, employment status, home addresses, and in some cases information involving spouses and Social Security numbers. An FBI spokesperson confirmed the bureau is aware of unauthorized-activity claims involving FBIJobs.gov and is investigating.
Axios said it could not independently verify whether all of the allegedly stolen data is authentic or current, although cybersecurity researchers said the incident appears credible, and 404 Media obtained a sample reportedly containing information on roughly 5,000 alleged agents. ShinyHunters also claims it accessed other FBI services and exploited a zero-day affecting Oracle PeopleSoft to compromise the jobs site. Those broader claims remain under investigation. If verified, the breach could create unusually serious risks because personnel information can be used for phishing, coercion, identity theft, and targeting of employees’ families.
Why It Matters: Breaches involving government personnel can create national-security and personal-safety risks far beyond the normal consequences of stolen corporate records.
Source: Axios.
Palo Alto Networks brings OpenAI and Anthropic models into continuous AI cyber defense
Palo Alto Networks has launched Unit 42 Continuous Frontier AI Defense, a new service that uses advanced AI models from Anthropic and OpenAI alongside other models to continuously probe corporate systems for security weaknesses. The service is intended to examine web applications, APIs, cloud infrastructure, and possible attack paths, then help organizations prioritize and remediate vulnerabilities before attackers exploit them.
The product is another sign that cybersecurity is becoming one of the first enterprise markets where AI agents may actively compete against other AI systems. Palo Alto says attackers are already using automation to compress operations that once took weeks into hours. The company argues that no individual AI model performed well enough across every vulnerability class in testing, which explains its multi-model approach. More broadly, cybersecurity vendors increasingly face a race to automate both offensive testing and defensive response as autonomous agents gain access to corporate networks, credentials, APIs, and business applications.
Why It Matters: AI is shifting cybersecurity from periodic, human-led testing to continuous, machine-speed attack simulation and remediation.
Source: Reuters.
OpenAI Hires Patreon Co-Founder to Build Creator AI Tools
OpenAI has hired Patreon co-founder and longtime technology chief Sam Yam to lead a new Creator Product group, and he is bringing former Patreon product head Drew Rowny and engineering head Shannon Ma with him. Yam said on X that the team will “build together with Creators” and offer early access to tools he described as “critically valuable to Creators and their communities,” pointing to OpenAI DevDay on September 29 and to a separate announcement Sam Altman has teased for this week. The move gives OpenAI the operators who spent 13 years turning fan subscriptions into a default business model for independent media.
The hire lands as OpenAI tries to turn ChatGPT from a chatbot into a platform that can take payments, run agents, and keep users inside its own ecosystem. Creator tools would sit next to existing product bets in health, commerce, and coding, and they would compete with YouTube, TikTok, and Adobe for the layer where AI-generated video, voice, and agents meet monetization. If DevDay shows subscription, tipping, or licensing rails baked into ChatGPT, OpenAI would be using distribution it already owns rather than renting audience from app stores and social networks. That directly challenges platforms whose economics depend on being the checkout page for digital work.
Qualcomm Splits Its First 2nm Flagship Into Two AI Phone Chips
Qualcomm unveiled the Snapdragon 8 Elite Gen 6 and Snapdragon 8 Elite Extreme Gen 6, its first flagship mobile processors built on TSMC’s 2-nanometer node and, Qualcomm says, the first to run CPU prime cores at 5 GHz. Both chips use an eight-core Oryon design with two 5 GHz prime cores and six 4 GHz performance cores, plus a reworked Adreno GPU Qualcomm rates about 35 percent faster and 40 percent more efficient than last year’s Elite Gen 5. The Hexagon NPU adds an Element Accelerator for transformer workloads. The standard Elite is billed as 10 percent faster on CPU and 37 percent more power-efficient; the Extreme variant adds a 50 percent larger shared memory pool, dedicated Adreno matrix cores, and on-device support for mixture-of-experts models above 30 billion parameters.
The Extreme chip targets agentic features that stay on the handset: long-context assistants, 8K capture, and local models that don’t send every prompt to the cloud. Motorola, Xiaomi, and ZTE are among the first phone makers named, and Motorola previewed a Signature 27 handset on the Extreme silicon before year-end. Qualcomm also highlighted a sensing hub that can run models up to about 200 million parameters for speaker ID and background automation, plus new controls over what on-device agents can access. The split SKU responds to OEMs that want a true flagship and a slightly cheaper “flagship-enough” part as memory prices squeeze the smartphone bill of materials.
Why it matters: On-device agentic AI only scales if the NPU, memory, and power budget live in the phone — and a two-SKU 2nm launch is Qualcomm’s bid to keep that stack from migrating entirely to Apple and cloud APIs.
Microsoft disrupts EvilTokens AI phishing network linked to more than 12,000 compromised accounts
Microsoft has disrupted EvilTokens, a phishing-as-a-service platform that allegedly compromised more than 12,000 Microsoft accounts across more than 10,000 organizations. Unlike conventional phishing kits, EvilTokens incorporated AI throughout the attack process, including analyzing compromised inboxes, identifying valuable relationships and payment authorities, selecting targets, and drafting impersonation messages. Authorities in the United Kingdom also arrested two suspected administrators of the service.
That distinction is important. Generative AI is no longer limited to making scam emails sound more convincing. Platforms such as EvilTokens can potentially help criminals decide who inside an organization is worth impersonating, which financial relationships to exploit, and which messages are most likely to trigger a payment or credential handoff. That could make sophisticated business-email-compromise techniques available to attackers without years of social-engineering experience. Microsoft coordinated the operation with law enforcement, Health-ISAC, and identity intelligence provider SpyCloud.
Why It Matters: AI is lowering the expertise needed to run targeted fraud campaigns, turning cybercrime infrastructure itself into an increasingly automated software business.
Source: BleepingComputer.
Google Chrome 154 patches 108 security flaws, including 11 critical vulnerabilities
Google has released Chrome 154 for Windows, macOS, and Linux with fixes for 108 security vulnerabilities, including 11 rated critical. The patched issues include buffer overflows, out-of-bounds writes, and use-after-free vulnerabilities affecting components including ANGLE, WebGL, GPU handling, ServiceWorker, Fullscreen, WindowDialog, and AdFilter. Google said external researchers reported 32 of the flaws.
Google has not said that the Chrome 154 vulnerabilities are being actively exploited, but several involve memory-safety defects that can become serious attack vectors when chained with other weaknesses. Chrome’s enormous desktop footprint also means browser flaws can expose businesses and consumers at considerable scale. The update is rolling out as version 154.0.8037.57 for Linux and 154.0.8037.57/.58 for Windows and macOS. Organizations that manage large browser fleets face an increasingly short window between vulnerability disclosure, proof-of-concept development, and attempted exploitation.
Why It Matters: Browsers have become central enterprise computing platforms, making rapid patching increasingly important as attackers automate vulnerability discovery and exploitation.
Source: SecurityWeek.
Arista warns of actively exploited critical VeloCloud zero-day with maximum CVSS score
Arista Networks is urging customers to immediately patch a critical vulnerability affecting on-premises VeloCloud Orchestrator deployments after confirming that the flaw has been exploited as a zero-day. Tracked as CVE-2026-93952, the vulnerability has a CVSS severity score of 10 and could allow a remote attacker to access privileged internal functionality. Successful exploitation could compromise the confidentiality, integrity, and availability of both the orchestrator and the data it manages.
VeloCloud Orchestrator centrally configures, monitors, and manages SD-WAN infrastructure, making it a particularly valuable target. Compromising management infrastructure can give attackers leverage far beyond a single endpoint because these systems often sit at strategic control points across distributed corporate networks. Arista says the vulnerability was externally discovered and is known to be under active exploitation, raising the urgency considerably compared with a theoretical security bug.
Why It Matters: Attacks on network-management platforms can provide attackers with broad access across enterprise infrastructure, making actively exploited flaws especially dangerous.
Source: SecurityWeek.
Comma.ai’s hands-off driving tech faces federal investigation after fatal crashes
The U.S. National Highway Traffic Safety Administration has opened an investigation into aftermarket driver-assistance technology from Comma.ai following five reported crashes, including two incidents that resulted in three deaths. Regulators said the crashes involved vehicles equipped with Comma-related systems failing to detect or react appropriately to slow or stopped vehicles in the same lane. Four of the incidents reportedly involved as many as 11 injuries.
Comma.ai sells hardware that works with its openpilot software to provide steering, acceleration, braking, adaptive cruise control, and lane-centering features on compatible vehicles. The situation is complicated by the open ecosystem around openpilot: regulators said some incidents may have involved modified third-party forks rather than Comma.ai’s original software. That raises difficult safety and accountability questions as increasingly sophisticated driving software becomes customizable outside the tightly controlled environments used by traditional automakers. Comma warns drivers that they remain responsible and must be ready to retake control.
Why It Matters: The investigation tests how regulators will handle increasingly capable automotive software that can be installed — and modified — independently of vehicle manufacturers.
Source: TechCrunch.
AT&T says automation and AI will reshape its workforce as telecom giant cuts legacy infrastructure
AT&T expects to operate with fewer employees five years from now as the telecom giant uses automation and AI to streamline operations and retires portions of its legacy infrastructure. Chief Technology Officer Jeremy Legg told WIRED that the company does not expect to maintain its current headcount as it automates internal processes and continues moving away from older systems such as traditional landline infrastructure.
The comments provide a concrete example of how AI may affect employment inside large technology-adjacent companies. Rather than replacing a single profession overnight, automation is being layered into network operations, customer service, engineering workflows, and administrative functions while companies simultaneously modernize physical infrastructure. AT&T has already reduced its workforce significantly over the past decade. For enterprise technology vendors, telecom operators represent an important proving ground because they combine huge customer bases with massive networks, decades of legacy systems, and strong incentives to lower operating costs.
Why It Matters: AT&T shows how the first major labor impact of enterprise AI may come through gradual automation combined with broader infrastructure consolidation.
Source: WIRED.
Xiaomi releases trillion-parameter MiMo-V2.6 AI model family for reasoning and agents
Xiaomi has released its MiMo-V2.6 family of AI models, including MiMo-V2.6-Pro, MiMo-V2.6-Flash, and an ultra-speed version aimed at latency-sensitive applications. Xiaomi describes the Pro model as an omni-modal trillion-parameter reasoning system built for long-horizon tasks, research, cybersecurity, and complex professional workloads, while Flash targets lower-cost, high-volume usage. The company is also releasing tools and model components intended to support research into reinforcement learning and AI agents.
The announcement adds Xiaomi to the growing list of Chinese technology companies building large models rather than relying exclusively on outside AI providers. It also illustrates how the Chinese AI market is broadening beyond specialist labs such as DeepSeek and Moonshot into consumer-electronics companies with enormous device distribution. Xiaomi’s performance claims still need broader independent benchmarking, particularly for real production workloads. But a capable model stack combined with smartphones, vehicles, home devices, and other hardware could give Xiaomi multiple ways to deploy AI directly to users.
Why It Matters: AI competition is spreading from standalone model labs to hardware companies that can place their own models across phones, vehicles, and connected devices.
Source: Xiaomi.
Anaconda Biomed raises $56 million to advance stroke-treatment medical technology
Barcelona-based medical-device startup Anaconda Biomed has secured roughly $56 million in fresh financing to advance its endovascular technology for treating acute ischemic stroke. The company plans to use much of the capital to complete its pivotal Athena study of the ANA Funnel Catheter, a device used during thrombectomy procedures to assist doctors removing blood clots in patients suffering large-vessel-occlusion strokes.
The funding reflects continued investor interest in medical technologies where engineering can materially improve clinical outcomes without waiting for entirely new drug classes. Anaconda is also developing complementary neurovascular devices including stent retrievers, guidewires, and sheaths. Omega Funds led the financing, joined by Spain’s FOCO co-investment fund and existing investors including Asabys, Ysios Capital, and CDTI. Medical-device startups face longer validation and regulatory cycles than typical software companies, so capital is often closely tied to clinical milestones rather than simple customer growth.
Why It Matters: Venture capital is continuing to back technically difficult healthcare platforms where successful clinical trials can create significant barriers to competition.
Source: Fierce Biotech.
Connect Ventures raises $55 million first close for fifth fund as European VC leans into deeptech
European venture firm Connect Ventures has reached a $55 million first close for its fifth fund, according to Sifted, as the investor looks to back another generation of startups with an increasing emphasis on technically ambitious companies and deeptech founders. Managing partner Sitar Teli told the publication that investors need to follow where the strongest founders are choosing to build.
The fundraise arrives during a notable shift in European venture capital. Generative AI remains dominant, but investors are also moving further into semiconductors, robotics, defense technology, advanced materials, climate infrastructure, and other sectors requiring more than a lightweight software product. Europe has deep technical talent and major universities but has historically struggled to translate that research into companies that can scale globally. New early-stage capital aimed at deeptech could help narrow that gap, although capital intensity remains a challenge compared with traditional SaaS investing.
Why It Matters: European venture firms are allocating more capital toward defensible technologies built around engineering, science, hardware, and proprietary know-how.
Source: Sifted.
Azulene Labs raises $3.4 million to build physics-based AI for drugs and advanced materials
Azulene Labs has raised $3.4 million in pre-seed funding to develop physics-based AI models for drug discovery, materials science, and industrial chemistry. Ground State Ventures led the round with participation from existing investor Entrada Ventures and angel investors. The startup says its systems train on quantum-mechanical data to make molecular simulations more accurate and computationally efficient.
The company sits at the intersection of two increasingly important areas: AI-assisted scientific discovery and computational chemistry. Traditional molecular simulation can demand substantial computing resources, while conventional machine-learning systems can produce faster predictions but may struggle when pushed outside their training distributions. Startups such as Azulene are betting that embedding more physical information into AI systems can narrow that gap. Success will ultimately depend less on headline model claims and more on whether the technology can consistently help researchers identify useful compounds, materials, or drug candidates faster than existing computational approaches.
Why It Matters: The next major AI market may include models trained around the laws of physics rather than primarily on human-generated text and images.
Source: The Quantum Insider.
U.S.-China “AI Hotline” Is a Phone Line Between Two Officials
Treasury Secretary Scott Bessent’s proposed U.S.-China “notification mechanism” for national-security AI incidents is, for now, an open line between Bessent and Chinese Vice Premier He Lifeng rather than a standing technical body, four people familiar with the arrangement told Politico. Chinese officials have not publicly accepted the channel. Bessent described it on CNBC as “a communications line, an incident line” and said the two governments had formalized “USA-China AI dialogues,” with another meeting expected in Shenzhen in about two months. The idea surfaced after Sunday talks in New York meant to tee up this week’s Trump-Xi meetings. One official briefed on the plan called the mechanism’s scientific term “BS.”
The thin design matters because both capitals are under pressure to show they can talk about runaway agents, cyber incidents, and biological misuse without accepting the other side’s preferred rules. Washington does not want a UN regulator; Beijing does not want U.S. export controls recast as safety policy. A two-person phone tree will not stop a model from being fine-tuned into a worm, but it is the first on-the-record attempt by the two largest AI powers to name a crisis contact. Startups watching export licenses, outbound investment screens, and cloud-region approvals should treat Shenzhen as the next calendar date that can move those rules, not as a standards conference.
Why it matters: Great-power AI safety is starting as diplomacy between two men, which is fragile — but it is also the only channel that can change export and compute policy overnight.
Source: Politico.
Zapata launches Quantum Pilot to help companies identify commercially useful quantum-computing applications
Zapata Quantum has introduced early access to Quantum Pilot, a cloud-based, hardware-agnostic platform intended to help companies determine where quantum computing could produce measurable business value. The system is designed to evaluate potential quantum applications, map them against expected hardware progress, and help organizations prioritize projects rather than committing prematurely to a particular quantum architecture or vendor.
That addresses a persistent problem in quantum computing. Hardware capabilities are progressing, but most enterprises still lack a clear way to decide which business problems might justify quantum resources and when those machines could outperform conventional systems economically. The sector now includes competing approaches based on superconducting systems, trapped ions, neutral atoms, photonics, and other technologies, making long-term platform selection difficult. A hardware-neutral planning layer could become valuable if quantum systems start crossing commercially relevant thresholds. For now, however, most proposed business advantages remain highly workload-specific and require careful validation.
Why It Matters: Quantum computing is developing a software and planning layer aimed at bridging the gap between experimental hardware and real enterprise adoption.
Source: Zapata Quantum.
Peloton launches cheaper foldable treadmill and expands AI-powered coaching
Peloton has unveiled three new treadmills, including the Tread Flex, its first folding model and a lower-cost attempt to broaden the company’s addressable market. The Tread Flex is priced at $2,195, while the new Tread Vision costs $3,495 and the Tread+ Vision is priced at $6,695. The Flex folds to reduce its footprint and targets consumers living in smaller homes or apartments. The new machines are scheduled to go on sale in the U.S. beginning October 1.
Peloton is also expanding Peloton IQ, its AI-assisted coaching software. New capabilities analyze metrics such as cadence, stride length, and vertical oscillation and can generate personalized pace recommendations. The combination reflects a broader trend in consumer hardware: manufacturers increasingly need software and AI services to differentiate devices that would otherwise compete primarily on physical specifications and price. For Peloton, the strategy matters especially after its pandemic-era growth reversed and subscriber momentum weakened.
Why It Matters: Consumer-device companies are increasingly using AI-driven personalization as a software layer to make mature hardware categories feel new again.
Source: Axios.
That’s your quick tech briefing for today. Follow us on X @TheTechStartups for more real-time updates.
