Artificial Intelligence - Page 4

ChatGPT accounts are getting a serious security upgrade, with users now able to lock access behind a physical USB key in a move that reflects just how valuable AI accounts have become.

The new feature, part of OpenAI's Advanced Account Security rollout, allows users to authenticate using hardware security keys instead of traditional passwords. These keys, which plug into a device via USB or connect wirelessly, act as a physical layer of protection, ensuring only someone with the device can access the account.

ChatGPT is no longer just a casual tool, for many users it now stores sensitive conversations, work documents, personal data, ideas, hopes, and even in some cases, extremely sensitive personal information. That makes it an increasingly attractive target for hackers, particularly through phishing attacks that exploit weak passwords. Hardware keys are widely considered one of the most secure authentication methods because they rely on encrypted credentials stored directly on the device rather than something that can be stolen or guessed.

A bank CEO has admitted to using an AI clone to host a conference call, and no one realized it wasn't him until he revealed it almost an hour into the meeting.

The incident, revealed in a recent interview, has the tech and corporate worlds buzzing, which was more than likely the intention of the stunt. The CEO of Customers Bank, Sam Sidhu, described the AI system as a real-time voice and language model trained to mimic his speech and decision-making, even handling follow-up questions with consistency and nuance.

He framed the stunt as a test of AI's capabilities, calling it "a bit of a prank" but also a demonstration of how automation can streamline high-level communication. The AI version of the CEO performed so convincingly that investors and stakeholders never suspected it wasn't the real person on the line. Notably, Customers Bank signed a deal with OpenAI last week, which enabled the AI company to integrate its AI agents into the banking giant.

With strict restrictions on what US-based hardware can be sold in China for GPUs and other AI-related technologies, the direct result, according to NVIDIA CEO Jensen Huang, is that the company's market share in the region has dropped to zero.

"In China, we have now dropped to zero," Jensen Huang said in a recent interview with the Special Competitive Studies Project. Naturally, with the boom in AI accelerator and data center equipment sales, a lot of money and market share are being left on the table. And yeah, Jensen Huang isn't happy with the situation, adding that it doesn't make "strategic sense" and that it has "already backfired" as Chinese-based technologies fill the void.

"Conceding an entire market the size of China probably does not make a lot of strategic sense, so I think that has already largely backfired," Jensen Huang continues. "Maybe it made sense at the time, but I think the policy really needs to be dynamic and needs to stay with the times. I think it would be fairly safe to say that having American chip companies and other companies in China makes a lot of sense."

YouTube is apparently working to integrate more AI features into its user interface. Shortly after increasing the price of YouTube Premium, they launched AI avatars for Shorts creators and began asking users whether they felt their feeds were filled with "AI slop". However, presumably after getting the answer that we weren't getting enough AI on our feeds, they are reportedly doubling down on AI features.

A new AI-powered search feature called "Ask YouTube" is reportedly in the works. Users will now get a tool in the search bar that uses AI to find the perfect answers to their queries, integrating videos into the answers. If you searched for a "how-to" style video, you might get a step-by-step instructional answer with multiple videos curated by AI to best answer your question.

In the demo, a user asks YouTube to "plan a 3-day road trip from San Francisco to Santa Barbara". The feature then proceeds to do its "AI magic" and provides a result with a mix of text, short videos, and long-form videos that are tailored to the user's specific query. This might seem like a red flag for discoverability, but YouTube claims it will show videos and relevant segments, along with channel details, to point users to the sources. You can then ask follow-up questions to the original query and get more video suggestions, much like the conversations we are now used to holding with large language models.

NVIDIA is pushing into the next phase of artificial intelligence with OpenClaw agents, signaling a shift from passive AI tools toward systems that can actively perform tasks.

Unlike traditional AI models that generate responses, OpenClaw agents are designed to execute workflows, make decisions, and operate across systems with minimal to no human input. Picture a virtual assistant that controls your entire machine, can take instructions, and carries out tasks. NVIDIA positions this as a major evolution in how organizations will use AI, particularly in enterprise environments where automation and efficiency are critical.

AI is moving beyond chat interfaces and into agent-based systems capable of handling complex operations across multiple platforms. NVIDIA even went as far as to call OpenClaw agents the ChatGPT moment for AI agents. Companies are increasingly looking for solutions that do not just assist users but actively complete tasks, reducing manual effort. NVIDIA's involvement suggests this shift could accelerate quickly, especially given its strong position in AI hardware and infrastructure.

Intel and AMD have partnered to standardize a new x86 ISA extension to supercharge AI performance. Under the x86 Ecosystem Advisory Group (EAG), AI Compute Extensions (ACE) are engineered to deliver a 16x increase in compute density over standard AVX10 instructions.

Traditional vectorized hardware (SIMD) and instructions, such as AVX, operate on one-dimensional data. Machine learning, powered by Matrix Multiplication and Accumulate (MMA), has created a demand for specialized hardware that runs in two dimensions. This need led to the rise of specialized Tensor Cores in GPUs, and now ACE brings that same logic to the x86 CPU.

ACE uses an outer-product algorithm and eight new 2D Tile Registers, each with 16x16 dimensions and 32-bit precision. Large AI datasets are split into sub-matrices, with the hardware consuming two 16x4 input matrices (at 8-bit precision) - the Blue and Yellow matrices. At each intersection on the 16x16 grid, the hardware computes an inner product between a 1x4 and a 4x1 vector.

Google Photos is evolving beyond photo storage with an AI-powered wardrobe feature that lets users catalog, mix, and virtually try on outfits.

The tool analyzes past photos to identify clothing items, building a digital wardrobe users can browse and combine. Whether you're reimagining a past outfit or experimenting with screenshots of apparel from social media or shopping sites, the feature overlays selections onto your body using augmented reality (AR).

The virtual try-on functionality leverages pose estimation and background removal to align clothes naturally with the user's body. The idea is to reduce the annoying amount of time spent trying on different outfits before leaving the house. The idea is that users now won't have to hunt through physical closets or manually search for matching pieces; they can simply try on what they have in their virtual dressing room.

Well, it's official: Microsoft and OpenAI are no longer monogamous when it comes to AI, with both companies releasing statements confirming that their exclusivity agreement is over. Prior to the announcement, Microsoft and OpenAI's several-year-long partnership saw the former delivering the Azure cloud-based infrastructure and billions of dollars in funding for the company's breakout AI hit, ChatGPT.

With the new agreement, "OpenAI can now serve all its products to customers across any cloud provider," however, Microsoft will remain OpenAI's primary cloud partner, with the AI firm's products to "ship first on Azure." This new open relationship makes sense in light of the recent news of Amazon investing $50 billion in OpenAI. A deal that includes silicon and gigawatts of compute.

In fact, when news broke that OpenAI was in talks with Amazon, some reports suggested that Microsoft considered suing OpenAI. A rumor that makes it sound like the partnership between the two companies has soured in recent months. That said, the revised agreement sees Microsoft retain its license for OpenAI's models and various products through 2032; however, this is now non-exclusive.

Meta continues to bet heavily on Agentic AI, as the company has signed an agreement with Amazon to deploy AWS Graviton processors at scale. In late February, Meta signed a deal with AMD to deploy 6 gigawatts of AI hardware. In March, Arm released the Arm AGI CPU for Agentic AI, built in collaboration with Meta. The new partnership with AWS is another step in the same direction for the company.

Recently, we have been seeing a significant shift in the compute demands for Agentic AI workloads. While GPUs remain the dominant force behind most AI workloads, CPUs are slowly gaining significance, thereby increasing demand. This is why Meta is now partnering with Amazon's AWS to deploy tens of millions of Graviton cores to its compute portfolio.

The Graviton5 chip is a 192-core chip built on the AWS Nitro System and features a cache that is five times larger than its previous generation. AWS claims an "up to 33%" improvement in core-to-core latency, greater bandwidth, and the ability for Meta to run its own virtual machines without compromises to performance or security.

NVIDIA has rolled out OpenAI's latest frontier model across its global workforce, with CEO Jensen Huang calling it a milestone in the "age of AI." Over 10,000 NVIDIA employees in engineering, product, legal, finance, and marketing were given early access to Codex, an agentic coding tool powered by GPT-5.5.

Engineers have been using the GPT-5.5-powered Codex for several weeks now, reporting big efficiency gains in software development and maintenance. "Debugging cycles that once stretched across days are closing in hours," says NVIDIA. Teams are also using natural-language prompts to deliver end-to-end features more reliably and with fewer wasted cycles than prior models.

Huang encouraged employees to adopt Codex in an internal email, describing AI agents as teammates boosting productivity across all roles, not just engineering. "Chatbots answer questions. Agents do work," Huang wrote. "Let's jump to lightspeed. Welcome to the age of AI." Employees have echoed that sentiment, describing Codex as both "mind-blowing" and "life-changing."

At Cloud Next 2026, Google introduced Workspace Intelligence, a new semantic layer for Google Workspace that creates a unified context by mapping emails, chats, files, collaborators, and active projects. The idea is to shift Workspace from a collection of productivity apps into an AI-driven system that can replicate how you actually work.

Workspace Intelligence is aimed at organizations that want AI to be more than just a chatbot. Google Workspace head of product Yulie Kwon Kim says the system is intended to "emulate the institutional knowledge of a long-tenured employee, but for AI." Google says it can gather material, rank priorities, track stakeholders, and adjust outputs to user writing styles.

To build a truly useful AI coworker, you need context, and a lot of it. Workspace Intelligence's goal is to provide rich organizational context from a company's documents, emails, presentations, and messages. This lets the system do things like generate content in your voice, picking up on preferences like whether you never use exclamation marks or love using them.

A new Trojan dubbed "OpenClaw" is raising serious alarms, with researchers warning that AI agents are now being weaponized to take full control of thousands of systems.

Security analysts report that OpenClaw has already compromised more than 28,000 machines, leveraging AI-driven automation to execute commands, adapt to environments, and maintain on these systems in ways that traditional malware struggles to achieve. The key concern here isn't the scale or the number of affected machines, but the infection's capability. For example, OpenClaw effectively hands attackers a semi-autonomous operator inside a system, and this operator has access to the entire machine.

According to a TechRadar report, the malware uses these AI agents to dynamically interact with compromised environments. Because the AI agents have access to the entire machine, the malware can make real-time decisions and control the system. This means attackers can automate system monitoring, perform lateral movement across the layers of the system they have access to, and even conduct data extraction.

Anthropic's unreleased Claude Mythos is at the center of a cybersecurity debate. Its coding capabilities are so powerful that in pre-release tests, the model identified thousands of previously unknown vulnerabilities in major operating systems and web browsers. Since then, many have questioned whether the model is truly the future of cybersecurity or simply a normal step forward wrapped in overhype.

Mozilla has now tipped the scales in Anthropic's favor, sharing in the zero-days are numbered blog post that early access to Mythos Preview helped it pre-identify 271 security vulnerabilities in the latest Firefox release. Firefox CTO Bobby Holley added that "in the never-ending battle between cyberattackers and cyberdefenders, defenders finally have a chance to win, decisively."

According to the post, Mythos detected these hundreds of vulnerabilities simply by analyzing unreleased code from the browser's latest version. The post didn't go into detail on the severity of these vulnerabilities, but Holley compared Mythos to earlier AI tools, noting that Anthropic's Claude Opus 4.6 had detected only 22 security bugs when analyzing Firefox 148.

A new and potentially explosive report from Reuters claims that Meta is installing tracking software on all its employees' computers to log mouse movements and keyboard keystrokes. According to the report, this is part of a new "broader initiative" to overhaul the company's workforce by training and building AI agents that can "perform work tasks autonomously."

The report notes that it has seen the internal memos from Meta regarding this, which is being called the Model Capability Initiative (MCI). Apparently, it will run on all work-related apps and websites, and will even take screenshots of employees' displays. According to one of the internal memos, the purpose of the new initiative is to accelerate AI model training in areas where they struggle.

This includes choosing options from dropdown menus and using keyboard shortcuts. "If we're building agents to help people complete everyday tasks using computers, our models need real examples of how people actually use them, things like mouse movements, clicking buttons, and navigating dropdown menus," Meta spokesperson Andy Stone said.

TSMC's CapEx (Capital Expenditures) for 2026 is projected to hit $56 billion. This capital is being funneled into a three-pronged effort to address rising chip demand: new fabrication plants, increased capacity, and a massive build-out of supporting infrastructure. However, despite this record-breaking spending, TSMC CEO C.C. Wei admits that shortages will likely persist until 2027 and potentially beyond.

The primary driver of this insatiable demand is, as you guessed, AI hardware, including GPUs (Nvidia/AMD), CPUs, and HBM (High Bandwidth Memory). This demand from NVIDIA, AMD, Intel, and even automotive giants is drying up capacity for TSMC's mature 3nm process node family, which, despite already reaching production targets, is still not enough.

The next issue pertains to packaging constraints. Modern AI processors aren't monolithic chips. Instead, they use technologies such as CoWoS (Chip-on-Wafer-on-Substrate) to integrate logic with HBM (High Bandwidth Memory). As it stands, TSMC's advanced packaging capacity is significantly tighter than its front-end wafer fabrication capacity. Another constraint is lead times, as each new fab needs years to reach high-volume production.

Google could be preparing a major shake-up in the AI hardware space, with reports suggesting the company is working on new custom silicon designed to challenge NVIDIA's dominance.

According to a Reuters report, Google is in talks with Marvell Technology to develop two new chips to improve AI model efficiency. The first is a memory processing unit (MPU) designed to work alongside Google's existing Tensor Processing Units (TPUs), while the second is a next-generation TPU specifically optimized for AI inference workloads.

The MPU is particularly interesting because it represents a shift toward offloading memory-intensive tasks from the main accelerator, something that has become a major bottleneck in modern AI systems. By separating data movement and memory operations, Google could significantly improve performance and reduce latency.

Graphic designers and casual users are still figuring out Canva's new AI 2.0, which can generate full designs from simple text prompts, and Anthropic is already showing off its own take. The company announced on Friday that it's launching Claude Design, a new experimental product that allows subscribers to use Claude to generate designs, prototypes, slides, and more.

As you might expect, users describe what they want, and Claude will create an initial version. From there, users can directly edit or ask the chatbot to further fine-tune. Claude will also generate custom sliders that users can push and pull to modify corresponding elements. In the demo video, Claude lets the user adjust the glow and density of arcs it uses to illustrate a connected network.

Claude Design can also apply a team's design system to every project it creates by reading a company's codebase and design files. "Every project after that uses your colors, typography, and comments automatically," according to the company.

Google recently launched an AI tool for creating designs from descriptions. Now, Canva is going down a similar route. The company has announced Canva AI 2.0, a more powerful version of its AI assistant that it claims will fundamentally change how people think about design work.

Canva AI 2.0 leans into the "vibe coding" idea but applies it to graphic design, letting users create fully editable designs from text prompts. Instead of starting with a blank canvas and building everything manually in a linear fashion, you just describe what you want, and the assistant calls the required tools and generates a few options to choose from.

Whatever it generates is a layered, editable output that can be refined with manual tools or by continuing to chat with the assistant. Every visual element it generates sits on its own layer, giving you the flexibility to tweak anything individually. You can also generate content across multiple formats in one go, perfect for anyone running multi-channel campaigns.

Researchers at the University of Washington have unveiled Vuebuds, a prototype pair of wireless earbuds equipped with tiny cameras that enable real-time visual AI interactions.

Designed to change how users engage with their surroundings, Vuebuds allow wearers to simply look at an object and ask questions, with responses delivered through the earbuds in about a second. Unlike smart glasses, which have struggled to gain mainstream adoption, Vuebuds aim to provide a more subtle and intuitive alternative.

The system captures still images using low-resolution, black-and-white cameras rather than recording continuous video, a deliberate choice to address privacy concerns while still delivering useful contextual information. A key feature of Vuebuds is on-device processing, meaning image analysis happens locally instead of being sent to the cloud.

Microsoft's AI-powered Recall feature is once again facing privacy and security concerns, despite a major redesign intended to address earlier backlash. Originally criticized as a "privacy nightmare," Recall captures snapshots of user activity on Windows PCs, storing everything from on-screen text to messages, documents, and browsing history.

After delaying the feature for nearly a year, Microsoft introduced stronger protections, including a secure data vault, Windows Hello authentication, and a Virtualization-Based Security (VBS) enclave. The company claimed these measures would prevent malware from accessing Recall data, even if it attempted to exploit user authentication, while simultaneously reassured users that Microsoft isn't going to be accessing the snapshots taken of desktops as they are stored locally.

However, cybersecurity researcher Alexander Hagenah has challenged those claims with a new tool called TotalRecall Reloaded. The tool can reportedly trigger a Windows Hello prompt and, once the user authenticates, extract all stored Recall data. Hagenah argues this shows malware can still "ride along" with legitimate authentication - something Microsoft said its redesign would block.