<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by CSI MJCET on Medium]]></title>
        <description><![CDATA[Stories by CSI MJCET on Medium]]></description>
        <link>https://medium.com/@csi_mjcet?source=rss-550736826920------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*v4_co2C5ZMOgYQcQABGd9Q.png</url>
            <title>Stories by CSI MJCET on Medium</title>
            <link>https://medium.com/@csi_mjcet?source=rss-550736826920------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 17:41:57 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@csi_mjcet/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Beyond the frame : Turning 2D images into 3D worlds]]></title>
            <link>https://medium.com/@csi_mjcet/beyond-the-frame-turning-2d-images-into-3d-worlds-412b77943168?source=rss-550736826920------2</link>
            <guid isPermaLink="false">https://medium.com/p/412b77943168</guid>
            <category><![CDATA[3d-graphics]]></category>
            <category><![CDATA[gaussian-splatting]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[CSI MJCET]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 15:15:30 GMT</pubDate>
            <atom:updated>2026-11-19T15:15:30.923Z</atom:updated>
            <content:encoded><![CDATA[<p>Author: <a href="https://medium.com/u/96e6a250c68b">SYED MUKHTAR-11</a></p><p><strong>INTRODUCTION:<br></strong>Imagine capturing a photograph of a beautiful place but being able to experience it only through a flat image, without ever exploring what lies beyond the frame. What if you could step inside that photograph, move around, explore hidden perspectives and experience the scene from entirely different angles? What once existed as a collection of pixels could become an explorable 3D world. This fascinating possibility is what modern computer vision is gradually bringing closer to reality.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/1*ANq5j9oF3c4CqjRuKODyVg.png" /></figure><p>Behind this transformation lies a fundamental problem in computer vision: How can machines recover the 3D structure of the real world from 2D visual information?</p><p>The idea of reconstructing 3D environments from 2D images emerged from the need to help computers understand and interpret the physical world beyond flat visual representations. As photography, digital imaging and computer vision evolved, researchers began exploring ways to recover depth, shape and spatial information from two-dimensional images. One of the earliest milestones in this direction was Larry Roberts’ work in 1963, which demonstrated how computers could use geometric information extracted from 2D images to reconstruct simple three-dimensional polyhedral structures.</p><p>Over the following decades, these foundational ideas paved the way for increasingly sophisticated 3D reconstruction techniques. Traditional approaches such as Stereo Vision, Photogrammetry, Structure-from-Motion (SfM) and Multi-View Stereo (MVS) relied primarily on geometric principles to recover 3D information from multiple images. Although these methods made significant progress, creating highly detailed and realistic digital environments remained challenging, particularly when dealing with complex scenes, intricate textures and computational requirements.</p><p>Meanwhile, the growing demand for realistic digital environments in gaming, films, robotics, autonomous systems and immersive technologies encouraged researchers to explore more advanced approaches. This gradually led to the emergence of learning-based techniques and neural rendering, with Neural Radiance Fields (NeRF) marking an important milestone in this evolution. Building upon these developments, 3D Gaussian Splatting introduced a new approach to representing and rendering realistic 3D environments with remarkable speed.</p><h3><strong>Understanding 3D Reconstruction:</strong></h3><p>Reconstructing the three-dimensional structure of a scene is a classic and fundamental problem in computer vision. For humans, it is relatively easy to <strong>infer</strong> the three-dimensional structure of a scene from a 2D photograph. However, reconstructing 3D geometry from a single image is a complex problem unless some prior information is given.</p><p>The 3D reconstruction allows you to recover the 3D structure from 2D images.</p><p>For example, classical algorithms approach to 3D reconstruction using multiple images and geometric constraints to 3D structures whereas other algorithms use is to get information about the 3D environment to conclude the 3D structure.</p><p>Structure from Motion</p><p>Structure from Motion (SfM) is one of the methods for 3D Reconstruction that allows us to reconstruct <strong>3D structures</strong> (points in space) by estimating <strong>camera motion</strong> (poses and intrinsics) using only a sequence of 2D images captured from different viewpoints. In other words, SfM takes ordinary photographs and concludes both where the camera moved and what the scene looks like in 3D without requiring depth sensors, GPS, and manual annotation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/1*wiTWY600x81LaP1dYcDq7g.png" /></figure><p>Some of the most common 3D reconstruction techniques include:</p><p>Binocular Stereo 3D Reconstruction: It mimics human depth perception by capturing a scene simultaneously from two calibrated cameras separated by a known horizontal distance. By matching corresponding pixels between the left and right images to calculate disparity, the system uses geometric triangulation to estimate depth. This foundational approach converts flat stereo pairs into dense 3D point clouds, making it essential for real-time applications like robotics and autonomous navigation.</p><p>Multi View Stereo Reconstruction: It is done by taking a reference image and a set of source images as input and using them to produce a depth map, voxels, mesh or another representation of the 3D structure of the image</p><p>Object Focused 3D Reconstruction: It uses the deep learning approach where the cameras are generally calibrated with respect to one another and focus on a single object of interest for 3D Reconstruction</p><p>SLAM based 3D Reconstruction: Simultaneous Localization and Mapping is a 3D reconstruction technique that builds a real time spatial map of an environment while simultaneously tracking the sensor or camera’s position.</p><h3>Understanding 3D Gaussian Splatting:</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/1*UMsNNL8pTf2VByYoViTVMA.png" /></figure><p>The early attempts however, were limited by their dependence on or heavy utilization of dense point clouds/dense grids leading to complex challenges.</p><p>Despite the advancements of SfM and NeRF, which provide improved reconstruction quality, these methods still face challenges related to computational cost and scene manipulation.</p><p>In this context, 3D Gaussian Splatting emerges not merely as an incremental approach but also as a game changing approach that redefines the boundaries of scene representation and rendering.</p><p>Gaussian refers to the Gaussian distribution, a bell-shaped normal distribution to represent the 3D scene and splatting refers to the process of projecting a 3D point into a 2D plane</p><p>In 3D Gaussian Splatting, a scene is represented as a set of 3D Gaussian primitives, each described by a mean position, covariance matrix, and opacity, each Gaussian primitive stores appearance information using spherical harmonic coefficients</p><p>It works mainly in four steps:</p><p><strong>Data Collection</strong>: The process begins with collecting multiple images or video frames of a real-world object or environment from different viewpoints. These images provide the visual information required to reconstruct the scene in 3D. Common datasets used for research and evaluation include Replica, ScanNet and specialized medical or surgical scene datasets.</p><p><strong>Processing:</strong> The collected images are processed to identify corresponding features and estimate camera positions using Structure-from-Motion (SfM). The resulting sparse 3D points are used to initialize 3D Gaussian primitives. Each Gaussian is defined by learnable parameters such as position (mean), shape, scale, orientation (covariance), opacity and color. These primitives collectively form the initial representation of the scene.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/931/1*TM3SmXyJpZHixYuHLNz4Kg.png" /></figure><p><strong>Optimization:</strong> The initialized Gaussian parameters are optimized by minimizing the difference between the rendered images and the original reference images. During this iterative process, parameters such as position, color, opacity and covariance are updated to improve reconstruction quality. Adaptive density control further refines the representation through operations such as cloning, splitting and pruning Gaussians. This allows the model to allocate more primitives to regions containing fine details while removing unnecessary ones.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/1*TFxa2O8VyWL6py--2TgNQA.png" /></figure><p><strong>Rendering:</strong> The final stage involves rendering the optimized Gaussian primitives using a differentiable rasterization technique. The 3D Gaussians are projected onto a 2D image plane, where they appear as 2D elliptical shapes. These projected Gaussians are sorted according to depth and combined using alpha compositing based on their opacity. This process produces realistic images from different viewpoints while supporting efficient rendering.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/1*p1ONZvfn5rI4YSY7ZYXdRw.png" /></figure><h3><strong>Applications:</strong></h3><p>In the production of Jurassic World: Rebirth (2025), 3D Gaussian Splatting transitioned from an academic breakthrough into a critical Hollywood pre-production tool. Working with visualization studio Proof, Inc., director Gareth Edwards utilized 3DGS to rapidly digitize real-world scouting locations — including dense jungle swamps, forest canopies, meeting spaces, and multiple coastlines — using simple iPhone and 360-camera captures.</p><p><strong>VR and XR: Meta’s Horizon Hyperscape Capture</strong> (2025) brings photorealistic 3D reconstruction directly to consumers. Using only the onboard cameras of a Meta Quest 3 or Quest 3S headset, users can walk through and scan any physical room in minutes. The captured data is processed in Meta’s cloud into a high-fidelity 3D Gaussian Splatting scene that preserves authentic lighting, shadows, and surface reflections.</p><p><strong>Real Estate and Architecture: </strong>The real estate industry has embraced 3D Gaussian Splatting because it solves a fundamental problem: showing spaces to remote buyers. It enables virtual property tours, architectural visualization, renovation previews, and preconstruction sales, making it easier for businesses to close deals faster, reducing costly on-site visits, and build buyer confidence from anywhere in the world.</p><h3><strong>Challenges:</strong></h3><p>A key challenge concerns the alignment between geometry and semantics. While semantic features can be associated with Gaussian primitives, ensuring consistent aggregation of semantics across viewpoints and scales remains difficult. In practice, errors in 2D supervision, imperfect projection and occlusions can introduce errors during 3D fusion leading to inconsistent semantics across views. Similar issues have been observed in natural rendering and feature fields, indicating the need for methods that explicitly enforce cross view semantic consistency. Overall, future research in 3D GAUSSIAN splatting is expected to focus on improving stability, robustness and generalization, while enabling richer interaction and more structured scene understanding.</p><p>These challenges highlight the need for integrated approaches that combine advances in 3D reconstruction, representation learning and multimodal reasoning</p><h3><strong>Future of Gaussian Splatting:</strong></h3><p><strong>4D Gaussian Splatting</strong>: It records not just static spaces but performances, events and changing environment. Early applications include volumetric video for concerts and sports, dynamic product showcases, and interactive storytelling.</p><p><strong>WebXR Integration:</strong> Gaussian splatting pairs naturally with virtual and augmented reality. Expect tighter integration with WebXR, enabling immersive experiences directly in browsers without native apps</p><h3><strong>Conclusion:</strong></h3><p>The journey from 2D images to 3D worlds represents one of the most remarkable advancements. What once seemed like an impossible task — recovering depth, shape and structure of the real world from photographs has evolved through techniques like 3D Gaussian splatting, NeRF, SfM, etc. Among these developments, 3D Gaussian splatting has emerged as a promising technique offering highly realistic scene representation and real time rendering capabilities.</p><p>Thus, 3D reconstruction is opening new possibilities for how machines perceive and interact with the physical world. Ultimately, the goal is no longer simply to view the world through images, but to recreate it in an interactive three-dimensional experience that can be explored, understood from different perspectives.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=412b77943168" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Global Microsoft 365 Outage Disrupts Outlook, Teams, and Cloud Services for Over 14 Hours!]]></title>
            <link>https://medium.com/@csi_mjcet/global-microsoft-365-outage-disrupts-outlook-teams-and-cloud-services-for-over-14-hours-7140e8a864d2?source=rss-550736826920------2</link>
            <guid isPermaLink="false">https://medium.com/p/7140e8a864d2</guid>
            <category><![CDATA[microsoft-outlook]]></category>
            <category><![CDATA[cloud-computing]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[microsoft-365]]></category>
            <category><![CDATA[distributed-systems]]></category>
            <dc:creator><![CDATA[CSI MJCET]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 14:20:43 GMT</pubDate>
            <atom:updated>2026-11-16T14:40:09.086Z</atom:updated>
            <content:encoded><![CDATA[<p>Author: <a href="https://medium.com/u/562c2b2b53e1">Nabiha Fatima</a></p><p>Imagine rushing to the airport with your bags packed, ready to catch an important flight. You check your phone for your boarding pass and the latest gate updates, only to find your inbox frozen. No alerts come through to tell you whether your flight is delayed, rescheduled, or boarding early. You try reaching out to the airline’s customer support, but their systems are down too, leaving you stranded at the terminal watching your plane take off without you, all because a cloud server failed miles away.</p><p>For millions of people on November 12, that exact frustration wasn’t just a bad dream, it was reality. <strong>A massive global Microsoft 365 outage knocked out Outlook, Teams, OneDrive, and Copilot for over 14 hours across 74 countries</strong>, with full recovery taking nearly 24 hours.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/633/1*HnSnOGdNiH9utSWOpqVXag.png" /></figure><p>As the backbone of modern corporate communication, Microsoft powers the daily email, calendar, and cloud infrastructure for tens of millions of businesses, airlines, and public institutions worldwide. When a flaw struck Exchange Online, the core cloud engine behind Microsoft’s email and scheduling tools, it triggered a chain reaction that stalled workflows across the globe.</p><h3><strong>How the Outage Disrupted Daily Life Around the World</strong></h3><p>The blackout quickly spread beyond corporate offices and hit critical healthcare services. In the UK and Canada, hospital staff lost access to digital communication and had to write patient records on paper. Pharmacies also faced delays in filling prescriptions because their online systems were down.</p><p>This disruption was not limited to just one region, spreading across <strong>74 countries</strong> and multiple continents from Asia-Pacific to Europe and the Americas.<br> Law enforcement and legal systems felt the strain as well. A judge in the UK had to send jurors home because court files would not load, while a sheriff’s office in the US asked people to pay bail in person after email payment alerts stopped working.</p><p>In India, the outage hit the aviation and corporate sectors the hardest. Major airlines like <strong>IndiGo</strong> and <strong>Akasa Air</strong> suffered disruptions to their customer email support systems, leaving many passengers waiting for replies. <strong>Air India</strong> also had to manually reroute its emails to keep passenger updates moving. Meanwhile, corporate offices across major Indian cities dealt with frozen inboxes and delayed meetings throughout the day, though central government departments escaped the disruption because their official emails run on a separate, independent network.</p><h3><strong>What Actually Caused the Crash</strong></h3><p>According to Microsoft, the outage started with a routine network capacity upgrade that went wrong. Engineers were rolling out an update meant to help their cloud systems handle heavier user traffic more smoothly. However, that update unknowingly carried a hidden software bug into the servers managing secure email connections. That single mistake quickly spiraled into a <strong>cascading failure</strong> across Microsoft’s data centers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/497/1*VMLxyBjBQdfcoiXXgR5jBw.png" /></figure><p>When the buggy update stressed the first batch of servers, automated safety checks stepped in and took those machines offline to protect the network. While that was meant to stop the problem, it actually made things worse, the sudden shutdown dumped a massive wave of user traffic onto the remaining healthy servers. Unable to carry the extra weight all at once, those servers began crashing one after another like falling dominoes, knocking Outlook, Teams, and other services offline around the world.</p><h3><strong>Was It a Cyberattack or Rogue AI?</strong></h3><p>As screens went dark across 74 countries, wild theories quickly started spreading online. People wondered if hackers had finally broken into Microsoft’s servers, or if a rogue AI tool had accidentally wiped out the network. Hacking rumors grew even louder because Microsoft had released an emergency security fix for Outlook just two days earlier. Meanwhile, many social media users guessed that Microsoft’s own Copilot AI had pushed a bad update by mistake.</p><p>However, both Microsoft and government cybersecurity officials quickly cleared the air, confirming that <strong>no hackers, cyberattacks, and AI tools were behind the crash</strong>. Experts explained that the earlier Outlook fix only patched a small issue on personal computers and had no link to Microsoft’s main cloud servers. Instead of a dramatic cyberattack or an AI gone wrong, the whole mess came down to a simple human mistake during a routine update.</p><h3><strong>What People Lost</strong></h3><p>While the outage felt like a total digital blackout, <strong>users did not permanently lose their emails, files, or calendars</strong>. Instead of deleting incoming messages, Microsoft’s servers safely held them in waiting queues until the network recovered. To make sure important emails did not bounce back to senders during the delay, Microsoft temporarily extended delivery windows to 48 hours, meaning delayed messages simply flooded into inboxes once systems came back online.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/608/1*HOxGiAG6Y-pDqjS0NAL54w.png" /></figure><p>However, the loss of time and money was massive. Millions of workers lost an entire day of productivity, missed important meetings, and struggled to serve customers. Economic analysts estimate that the global slowdown cost businesses over <strong>$7 billion</strong> in lost productivity. On top of that, Microsoft itself faces paying out up to <strong>$800 million</strong> in service-level refunds to paying corporate clients whose operations were stalled by the crash.</p><h3><strong>How Microsoft Resolved the Outage</strong></h3><p>Bringing the system back online took nearly 24 hours because engineers could not simply flip a switch and restart everything at once. Early attempts to reboot the servers failed when sudden waves of returning traffic overloaded them all over again. To break the cycle, Microsoft first isolated the failing machines so the crash would not spread any further, then rolled back and patched the faulty network update.</p><p>Once the bug was fixed, engineers carefully rerouted traffic and brought servers back online in small, controlled batches. They tested stability step by step starting with Asia-Pacific before moving on to Europe and North America, and cleared the remaining email backlogs by Saturday, November 14. While all services are now running normally again, the day-long blackout serves as a clear reminder of how much the modern world relies on a handful of cloud systems to keep daily life moving.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7140e8a864d2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Price of a Prompt: AI, Data Centres, and the Water We Don’t See]]></title>
            <link>https://medium.com/@csi_mjcet/the-price-of-a-prompt-ai-data-centres-and-the-water-we-dont-see-4ce792b01747?source=rss-550736826920------2</link>
            <guid isPermaLink="false">https://medium.com/p/4ce792b01747</guid>
            <category><![CDATA[sustainability]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[climate-change]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[data-center]]></category>
            <dc:creator><![CDATA[CSI MJCET]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 14:07:45 GMT</pubDate>
            <atom:updated>2026-11-12T14:18:56.280Z</atom:updated>
            <content:encoded><![CDATA[<p>Author: <a href="https://medium.com/u/e9342f866bb0">GazalaJabeenHafsa</a></p><p>Imagine sitting on your desk asking the AI to do your work while you enjoy your coffee. You probably didn’t realize that at the time a day’s worth of water is being used up in small towns. Why? Because their new neighbour is a place with no windows, no signs, no crowds. Just big grey buildings, one after another with power lines going towards them. These are Data Centers and they are being built faster than the world has ever seen. Behind each one is the reason: our growing need for artificial intelligence.</p><h3><em>Data Centers: what are they?</em></h3><p>Data centers are the hidden power plants of the internet: big secure buildings filled with rows of servers, storage systems and network devices that store, process and send the digital information that is part of daily life. Every time you watch a video send a message check the weather shop online or ask an AI a question data is likely being moved and worked on in one of these places. In short data centers turn electricity and equipment into the cloud services and online experiences that we often don’t think about.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/944/1*okSn2EMY9P1jYLlT93zLvw.png" /></figure><h3>The fast growth of the Data Centers</h3><p>Data centers have moved from corporate server rooms into some of the largest building projects in the world. The United States had 5,427 data centers at the end of 2025 and nearly 4,000 new U.S. Facilities have been announced, although only around 800 were being built, showing limits in land, equipment and especially electricity. The biggest change is size: an average facility may need megawatts but modern big data centers and AI areas now need 100 megawatts or more; the biggest projects being built could use 20 times more power than a normal AI data center.</p><p>This boom is being funded on a scale. Global data-center investment almost doubled from 2022 to 2024 reaching about $500 billion in 2024 while Amazon, Microsoft, Alphabet and Meta were expected to spend about $720–740 billion on investments in 2026 much of it going to AI systems. Growth is focused in Northern Virginia, Texas, Arizona, Ireland, Singapore and India’s Hyderabad, Mumbai and Chennai areas, where cloud providers look for power, strong internet connections and being close to users. The result is an increase in electricity use: from 415 TWh in 2024 to a forecasted 945 TWh by 2030 with AI as the main cause.</p><h3><em>The Fuel to this issue : Rise of AI</em></h3><p>The data-center boom is mostly driven by AI: chatbots, image and video creators and AI features now part of search, email, office software, coding tools and everyday apps. Each request starts inference, the process of running a model to get an answer so.</p><p>billions of questions add up to non-stop need. Inference is already than half of AI computing while training is still a big part as companies make bigger models.</p><p>AI is also being added to products that may not need it from spell checks and photo edits to “summaries and robot helpers. This creates an “ use” issue: easy defaults and free access push repeated creation even when a normal search or calculation would use less energy. Inference can use up to 65% of a model’s energy over its life making daily use a part of AI’s impact.</p><p>AI uses resources than regular computing because it uses tight groups of special chips, fast internet and water cooling; AI racks can use about 150 kW now with future designs going up to 1 MW.</p><h3><em>Water usage</em></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/941/1*hM26UxakwA_NoxL56VwBAw.png" /></figure><p>Servers, AI systems with lots of GPU create a lot of heat and need constant cooling to avoid problems and loss of performance. Many buildings use cooling: water takes in heat and then evaporates, which means the water is actually used up instead of just being taken from a river or water system</p><p>Water use has two parts. Direct use is cooling on the site; indirect use is the water used at power plants that provide the buildings electricity. In 2023 U.S. Data centers were estimated to use 17 billion gallons and 211 billion gallons indirectly through electricity generation</p><p>Estimates vary a lot depending on climate the number of chips and cooling methods so numbers per request should be taken carefully. A normal 100-megawatt big data center might use 2 million liters per day for cooling while large bases can need millions of gallons per day at the busiest times; Google’s plan for a Virginia base for example expected up to 8 million gallons per day. Hot weather can greatly increase use with evaporative systems using about 1–3 meters of water per megawatt-hour of computing</p><h3><em>THE WATER SHORTAGE PROBLEM</em></h3><p>Because cooling needs are highest in hot dry times data centers can fight with homes, farms and local water systems when water is least available. More than half of India’s data centers are in areas with water problems and the country’s four biggest centers Mumbai, Chennai, Hyderabad and Bengaluru have 75% of the national capacity.</p><p>The risks are clear in places. Chennai’s water sources almost ran out in 2019. Hyderabad is expected to have a 909-million-liter-per-day shortage for homes and businesses by 2027–28. In Virginia Google’s planned Botetourt County base first hoped</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/872/1*U4s469SBJE_boocIke5LNQ.png" /></figure><p>for 2 million gallons per day which could go up to 8 million gallons per day. The utility says the higher number is only for the hottest days.</p><p>Transparency is still an issue: utility companies and people often get little information about the water use of these places making long-term water planning hard. Because of this communities might only find out how water a project needs after its approved making it harder to change cooling systems or protect local supplies.</p><h3><em>Fixing the Water Problem</em></h3><p>The water issue isn’t unavoidable; it can be reduced with tech, smarter placement and more responsibility. Technologically data centers can use closed-loop cooling, which reuses water of evaporating it as well as water or immersion cooling that moves heat more efficiently. Air or waterless systems can stop use of cooling water though they might use more electricity. When water is needed operators can use treated waste water, recycled water or other non-drinkable sources of clean water. Waste heat can also be used for heating, greenhouses or industrial uses turning a problem into a benefit. Closed-loop and recycled-water systems are already being used by companies.</p><p>Smarter placement and operations are just as important. Facilities should be built where water is available electricity is clean and local areas can handle the demand. Operators can move work to times when cooling uses less water and energy. Site selection should match water availability, renewable energy supply and grid strength not just follow cheap land or tax breaks. Google for example looks at the health of a watershed. Uses air cooling or recycled water in areas with high water risks.</p><p>Laws must make these choices required. Governments should ask for reports on wateruse, water sources, cooling methods and energy use for each place. New projects should go through water risk checks, permits and limits in areas with water problems. Communities need to be involved before building starts and water prices should show scarcity of making clean water cheap. Mandatory reports and water-efficient cooling rules are already starting in some areas.</p><h3><em>Reducing the Demand Itself</em></h3><p>Technology can’t fix the issue if the need keeps growing. AI systems should be made to use power: smaller models where possible special models for specific jobs better ways to process and better hardware. A short answer should not need a general model.</p><p>Users and companies should also use AI carefully not by default. Not every email, image, search or routine decision needs an AI model; choosing the right tool for the job can stop extra processing.</p><p>However efficiency has a problem: the Jevons paradox. When AI gets cheaper and faster people and companies tend to use a lot more so savings per request can be less than the increase in requests. Efficiency is needed but not enough; limits on default use, avoidable automation and unnecessary tasks must go with it.</p><p>On the screen an AI answer shows up quickly clean and free.. Behind it are power plants, cooling systems, rivers and communities that may pay the real price. AI growth isn’t automatically the problem; unchecked and hidden growth is. The answer isn’t to stop progress. To make it clear and responsible: track water and power use, build where resources can handle it cool systems smartly. Use AI on purpose. Transparency, better design and local responsibility can keep both the technology and the areas, around it working. The question is no longer if AI will grow. If it will grow in a responsible way.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4ce792b01747" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Months of Planning, Hours of Theft: How a Fake Job Interview Led to the $6.34 Billion Binance Heist]]></title>
            <link>https://medium.com/@csi_mjcet/months-of-planning-hours-of-theft-how-a-fake-job-interview-led-to-the-6-34-billion-binance-heist-47392f1793c2?source=rss-550736826920------2</link>
            <guid isPermaLink="false">https://medium.com/p/47392f1793c2</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[crypotcurrency]]></category>
            <category><![CDATA[blockchain]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[ethical-hacking]]></category>
            <dc:creator><![CDATA[CSI MJCET]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 14:10:46 GMT</pubDate>
            <atom:updated>2026-11-09T14:10:46.652Z</atom:updated>
            <content:encoded><![CDATA[<p>Author: <a href="https://medium.com/u/df064f8b4fb7">Saad Riyaz Mohammed</a></p><p><em>How a fake job interview, compromised developer systems and a sophisticated wallet exploit led to one of the largest cryptocurrency thefts in history.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*G2T9OWIapnmN68G8I5YPHg.png" /></figure><p>Imagine a company managing billions of dollars in digital assets. Its infrastructure is built around cryptographic security, multiple layers of authentication and systems designed to prevent unauthorized transactions. Yet, within a matter of hours, billions of dollars begin moving out of its wallets without authorization.</p><p>This was the situation Binance faced on October 29, 2026.</p><p>One of the world’s largest cryptocurrency exchanges reported a security incident involving approximately $6.34 billion in digital assets. The attack affected multiple blockchain networks and exposed serious weaknesses in the systems responsible for managing and authorizing cryptocurrency transactions.</p><p>But the most interesting part of this incident was not simply how much was stolen. It was how the attackers got access in the first place.</p><p>According to Binance’s investigation, the operation involved compromised developer devices, malicious software, a fake recruitment process and unauthorized access to internal wallet infrastructure. What initially appeared to be an ordinary software installation eventually became part of a sophisticated attack against one of the largest players in the cryptocurrency industry.</p><p>To understand how this happened, we need to look at the events that unfolded months before the actual theft.</p><h3><strong>Understanding What Was Targeted</strong></h3><p>Before examining the attack, it is important to understand how cryptocurrency exchanges manage digital assets.</p><p>Exchanges like Binance generally maintain two types of wallets: hot wallets and cold wallets.</p><p>Hot wallets remain connected to the internet and are used to process everyday transactions such as deposits and withdrawals. This makes them essential for exchange operations, but their connectivity also increases their exposure to cyber threats.</p><p>Cold wallets are kept offline or in isolated environments. They are primarily used to store larger reserves of cryptocurrency with reduced exposure to remote attacks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*08Y9QpmBoSizUiLj2nR1ig.png" /></figure><p>The October incident involved Binance’s hot wallet infrastructure.</p><p>These wallets rely on internal systems that authorize transactions using cryptographic signatures. Binance also uses Multi-Party Computation (MPC), a cryptographic technique that distributes the signing process across multiple components rather than keeping an entire private key in one place.</p><p>This approach makes unauthorized access to private keys more difficult. However, it does not eliminate risks associated with compromised devices, credentials or transaction authorization systems.</p><p>And that is exactly where the attackers focused their efforts.</p><h3><strong>The Attack Began Months Earlier</strong></h3><p>According to Binance’s November 3 investigation report, the first identified stage of the attack dates back to June 2026.</p><p>An engineer working on Binance’s wallet infrastructure downloaded a tampered Homebrew installer while attempting to install an AI coding tool.</p><p>Homebrew is a popular package manager used by developers to install applications and development tools on macOS systems.</p><p>The compromised installer introduced malicious code into the engineer’s development environment.</p><p>Binance detected suspicious activity, replaced the affected device and rotated the associated credentials. However, the attackers had already gathered information about the engineer and other members of the wallet infrastructure team.</p><p>This information would become useful during the later stages of the operation.</p><p>The incident highlights a growing concern in software security: supply-chain attacks. Developers regularly depend on external packages, libraries and installation tools. When one of these trusted components is compromised, attackers can introduce malicious code into an otherwise legitimate development environment.</p><p>The initial compromise was only one part of a much larger operation.</p><h3><strong>A Fake Job Interview and a Compromised Developer</strong></h3><p>In September, another software engineer at Binance was approached through a legitimate job-hunting platform.</p><p>The individual claimed to represent another company and initiated what appeared to be a normal recruitment process. There were interviews, professional conversations and eventually a coding assignment.</p><p>Nothing seemed particularly unusual.</p><p>However, the assignment contained a malicious code repository. When the engineer downloaded and executed the project on a work computer, the malware compromised the development environment.</p><p>The malicious code was designed to communicate with an external server controlled by the attackers. It also used obfuscation techniques to make its activity harder to detect.</p><p>This was a sophisticated example of social engineering.</p><p>Instead of directly asking an employee for their credentials, the attackers created a situation in which the employee willingly downloaded and executed malicious software.</p><p>The technique is particularly concerning for technology companies because developers frequently work with unfamiliar repositories and third-party tools. A project that looks like a routine coding assignment can become an entry point into an organization’s internal systems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*c12MUVeAzv59ih9QD-R8_Q.png" /></figure><h3><strong>From a Compromised Device to Critical Infrastructure</strong></h3><p>The attackers did not stop after compromising the engineer’s computer.</p><p>Binance’s investigation describes how the attackers later used the compromised environment and an internal AI tool to gain access to the systems of two additional engineers.</p><p>During this process, they discovered a container deployment permission token stored on the computer of a wallet infrastructure engineer.</p><p>To understand why this mattered, consider how modern applications are deployed.</p><p>Organizations commonly use containers to package and run applications in controlled environments. Deployment credentials allow authorized users or services to introduce new containers into those environments.</p><p>If attackers obtain these credentials, they may be able to perform actions that appear to come from legitimate users.</p><p>Using the compromised credentials, the attackers accessed another engineer’s computer and deployed a container containing a malicious script.</p><p>This gave them a path into Binance’s wallet-signing infrastructure.</p><p>The attack had moved beyond individual employee devices. The attackers had reached systems involved in authorizing financial transactions.</p><h3><strong>The Technical Exploit: How Did the Attackers Move the Funds?</strong></h3><p>One of the most important technical elements of the incident involved Ethereum’s EIP-7702 authorization mechanism.</p><p>Introduced as part of Ethereum’s Pectra upgrade, EIP-7702 allows ordinary Ethereum wallets to delegate certain execution capabilities to smart contract code.</p><p>The feature was designed to make transactions more flexible and support improved wallet functionality.</p><p>However, Binance reported that the attackers used unauthorized access to its wallet-signing infrastructure to establish malicious EIP-7702 authorizations.</p><p>These authorizations allowed an attacker-controlled smart contract to gain partial execution rights over an affected hot wallet.</p><p>In simpler terms, the attackers found a way to use the wallet’s existing authorization mechanisms to execute operations that were not approved by Binance.</p><p>This is an important distinction. EIP-7702 itself was not a security vulnerability. The problem was that the attackers had gained access to the infrastructure responsible for authorizing these operations.</p><p>There was another complication.</p><p>Binance’s MPC-based wallet architecture made it difficult to immediately revoke the malicious permissions. Since the authorizations had already been established on the blockchain, simply suspending normal withdrawals could not invalidate them.</p><p>Binance had to develop and execute additional transactions to remove the unauthorized permissions.</p><p>This explains why the attack continued even after the exchange had started its containment process.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*ISgnegf2nbOdMVuyfCe4BQ.png" /></figure><h3><strong>Four Hours of Unauthorized Transfers</strong></h3><p>On October 29, the attackers began moving cryptocurrency from Binance’s hot wallets across multiple blockchain networks.</p><p>One of the earliest reported movements involved approximately 233.46 million XRP transferred from five Binance hot wallets within just four minutes.</p><p>The attack affected multiple cryptocurrencies across different blockchain networks, including approximately 233.46 million XRP, 970,500 SOL, 44.2 million OP and 18,270 BTC. These transfers highlight the scale of the breach and how the attackers targeted multiple digital asset networks rather than focusing on a single cryptocurrency.</p><p>Binance detected unusually large transactions and immediately began investigating. It suspended withdrawals and coordinated with external partners to freeze affected assets.</p><p>However, the existing on-chain authorizations meant that unauthorized transfers could continue.</p><p>According to Binance’s report, the final unauthorized transfer involved approximately 2,961.9 ETH and was finalized at 23:56 UTC.</p><p>The malicious EIP-7702 authorization was subsequently cleared at 00:10 UTC on October 30.</p><p>The incident demonstrated how quickly a compromise of internal financial infrastructure can turn into a large-scale loss, particularly when attackers gain access to systems responsible for signing transactions.</p><h3><strong>Who Was Behind the Attack?</strong></h3><p>On October 31, the Federal Bureau of Investigation attributed the theft to North Korea.</p><p>The FBI identified the activity as part of its TraderTraitor designation, which refers to North Korean malicious cyber operations targeting cryptocurrency and digital asset businesses.</p><p>Blockchain intelligence firm Chainalysis independently linked the incident to the same group after observing connections between funds stolen from Binance and assets associated with another North Korean-linked cryptocurrency theft.</p><p>The attribution is significant because North Korean cyber operations have repeatedly targeted cryptocurrency exchanges, blockchain platforms and technology companies.</p><p>These operations often involve sophisticated social engineering techniques and attempts to compromise employees who have access to valuable systems.</p><p>The Binance incident follows a similar pattern. Rather than relying entirely on a direct technical vulnerability, the attackers combined human deception with compromised software and internal access.</p><h3><strong>Following the Stolen Cryptocurrency</strong></h3><p>One interesting aspect of cryptocurrency investigations is that transactions are recorded on public blockchains.</p><p>Although wallet addresses do not directly reveal the identities of their owners, investigators can track how digital assets move between addresses.</p><p>Following the Binance incident, blockchain intelligence firms, law enforcement agencies and cryptocurrency companies began working together to trace and restrict the stolen funds.</p><p>According to Binance’s November 3 update, approximately $2.85 billion in digital assets had been frozen or intercepted.</p><p>This was made possible partly through cooperation with stablecoin issuers and other industry participants.</p><p>Some stablecoins, such as USDT and USDC, have centralized issuers that can freeze assets associated with certain addresses. Bitcoin and Ether do not have an equivalent central authority capable of arbitrarily freezing balances.</p><p>This creates different challenges when attempting to recover stolen cryptocurrency.</p><p>It is also important to distinguish between freezing assets and recovering them. Although a significant amount was restricted, this did not mean that all the stolen funds had been returned.</p><h3><strong>Binance’s Response and the Road to Recovery</strong></h3><p>Binance stated that it would absorb the financial impact of the incident rather than pass the losses on to its users.</p><p>In its November 3 report, the exchange said it had set aside more than 100% of the funds required to cover the loss. It also reported that 98% of the affected token assets had been replenished by October 31.</p><p>This replenishment came from Binance’s own resources. It should not be confused with recovering 98% of the cryptocurrency directly from the attackers.</p><p>The company also announced several security improvements, including rebuilding affected wallet infrastructure, replacing compromised devices, revoking exposed credentials and strengthening wallet-service controls.</p><p>It also committed to improving endpoint protection and AI-based behavioural anomaly detection.</p><p>These measures address several weaknesses identified during the investigation. However, their effectiveness will depend on implementation and independent security reviews.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*aZ3WzsoxHMIJ67vGNiiOPA.png" /></figure><h3><strong>What Does This Mean for Cybersecurity?</strong></h3><p>The Binance incident offers several important lessons that extend beyond cryptocurrency exchanges.</p><p><strong>Employee security is infrastructure security.</strong></p><p>The fake recruitment process demonstrates how attackers can exploit professional relationships to compromise otherwise secure organizations. Employees need to be cautious when downloading unfamiliar software or executing code from external sources.</p><p><strong>Developer environments require stronger protection.</strong></p><p>Development machines often have access to source code, internal services and deployment credentials. They should be treated as critical infrastructure rather than ordinary employee devices.</p><p><strong>Access permissions must be carefully controlled.</strong></p><p>The compromised deployment token illustrates the risks of excessive privileges. Organizations should follow the principle of least privilege and ensure that sensitive credentials cannot be easily accessed from compromised endpoints.</p><p><strong>Cryptographic security is only one part of the solution.</strong></p><p>MPC and secure wallet architectures can protect private key material. However, they cannot independently prevent compromised systems from initiating malicious signing operations. Transaction authorization requires additional safeguards and continuous monitoring.</p><p><strong>Incident response must consider blockchain-level permissions.</strong></p><p>Traditional controls such as suspending withdrawals may not be enough when malicious permissions have already been established on-chain. Exchanges must prepare for attacks that affect both internal infrastructure and blockchain execution mechanisms.</p><h3><strong>Beyond the Billions</strong></h3><p>The Binance incident is a reminder that even organizations with sophisticated security infrastructure are vulnerable when different parts of their systems become interconnected.</p><p>The attackers did not rely on a single weakness. They combined social engineering, software supply-chain compromise, stolen credentials and unauthorized transaction permissions to reach critical financial infrastructure.</p><p>The actual theft took place within hours, but the preparation behind it extended across several months.</p><p>For the cryptocurrency industry, the implications go far beyond one exchange or one security incident. Blockchain technology can provide transparency and cryptographic integrity, but the systems built around it still depend on people, software and access controls.</p><p>Ultimately, protecting digital assets requires securing the entire process through which they are accessed, authorized and transferred.</p><p>The Binance breach showed what can happen when even one part of that process is compromised.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*_jq_DmDf2QcSQiD-V04MCw.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=47392f1793c2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Google Loses $482 Billion in Market Value: Is AI the Beginning of the End for Traditional Search…]]></title>
            <link>https://medium.com/@csi_mjcet/google-loses-482-billion-in-market-value-is-ai-the-beginning-of-the-end-for-traditional-search-300f856eceef?source=rss-550736826920------2</link>
            <guid isPermaLink="false">https://medium.com/p/300f856eceef</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[stock-market]]></category>
            <dc:creator><![CDATA[CSI MJCET]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 13:08:55 GMT</pubDate>
            <atom:updated>2026-11-02T13:08:55.279Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Google Loses $482 Billion in Market Value: Is AI the Beginning of the End for Traditional Search Engines?</strong></h3><p>Author: <a href="https://medium.com/u/7018e4b2c8e6">Mohammed Nouman</a></p><h3><strong>1. Introduction to the Problem</strong></h3><p>Artificial Intelligence is changing the way people search for information and interact with technology. Instead of visiting multiple websites to find answers, users can now rely on AI platforms that provide quick answers, summaries, and recommendations. While this shift creates new opportunities for innovation, it also presents challenges for companies that depend on traditional ways of delivering information.</p><p>Alphabet, the parent company of Google, reportedly lost $482 billion in market value following its quarterly earnings report, as its shares experienced a significant decline. Despite strong growth in its cloud business, investors were concerned about rising AI investments, disappointing earnings, and the future of Google’s search advertising business. This raises an important question: Is Google’s investment in AI making the company stronger, or could AI eventually disrupt the very search engine that made Google successful?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/627/1*xx72S-BXel5TARrqoJTbyA.png" /></figure><h3><strong>2. Key Factors Leading to the Loss</strong></h3><p>One of the major reasons behind Google’s stock decline was its massive investment in Artificial Intelligence.</p><ul><li>Heavy AI Investment: Google is spending billions of dollars on data centres, advanced chips, servers, and AI infrastructure to compete with companies such as Microsoft and OpenAI. While these investments are necessary to remain competitive, developing and operating AI systems requires enormous financial resources, and the returns may take years to materialise.</li><li>Earnings Miss: Alphabet reported earnings per share of $2.12 against analysts’ expectations of $3.02, disappointing investors who had anticipated stronger financial results.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/654/1*cwt-PXC7B6jTx_OUMz8jrA.png" /></figure><ul><li>Negative Free Cash Flow: Heavy AI expenditure affected Google’s free cash flow, which represents the money remaining after accounting for operating expenses and investments. Investors are concerned that the company’s increasing investments could put pressure on the cash it generates.</li><li>Uncertain Returns: Investors are questioning whether Google can generate sufficient profits from its AI investments and whether its increasing expenditure will eventually translate into sustainable financial growth.</li></ul><h3><strong>3. Why the Loss Despite Growth in the Cloud Sector?</strong></h3><p>A key highlight of Google’s earnings report was the growth of its cloud computing business. Google Cloud recorded 103% revenue growth, indicating increasing demand for cloud infrastructure and AI services. Cloud computing allows companies to access computing power, storage, and software over the internet, and the growing adoption of AI has made this sector particularly promising.</p><p>However, strong revenue growth does not always translate into higher profits. Google must continue investing heavily in infrastructure, including data centres, servers, and advanced chips, to support its growing operations. These expenses increase overall spending and can put pressure on cash flow.</p><p>Another major concern is the possibility that AI-powered search could affect Google’s advertising business, which remains one of its primary sources of revenue. Traditional search encourages users to visit multiple websites through links, creating opportunities for advertisements. AI platforms, on the other hand, can provide direct answers without requiring users to visit several websites. This could reduce opportunities for traditional search advertising.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/739/1*fQTBtGXwgJx115utvhWxXw.jpeg" /></figure><p>As a result, despite the impressive growth in its cloud business, investors are looking beyond revenue figures and questioning whether Google can maintain profitability and generate sufficient returns from its investments.</p><h3><strong>4. Losses Faced by Investors and Market Reaction</strong></h3><p>Following the earnings announcement, Alphabet’s shares dropped by more than 12%, wiping out approximately $482 billion in market value. This does not mean Google lost that amount of cash. Instead, it means the market placed a lower value on the company.</p><p>When investors lose confidence in a company’s earnings potential and future profitability, they may begin selling their shares, causing the stock price and the company’s overall market value to decline.</p><p>The drop affected investors holding Alphabet shares, as the value of their investments decreased. Those who purchased shares at higher prices and sold during the decline faced losses, while investors who continued holding their shares were left uncertain about the company’s future performance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/626/1*Nue6o5FrBj1yWp1eHXTJXw.jpeg" /></figure><p>The market reaction highlighted growing concerns about technology companies, particularly whether spending billions of dollars on AI infrastructure and development can generate enough profit to justify the investment. It also demonstrated how quickly investor confidence can shift when expectations of future growth become uncertain.</p><h3><strong>6. What Does This Say About Google’s Future?</strong></h3><p>Google’s share price decline does not automatically mean that the company is losing its importance. The strong growth of Google Cloud demonstrates that the company continues to benefit from the increasing demand for AI and computing infrastructure. However, Google must adapt to changing user behaviour and ensure that its AI-powered search experience remains useful, competitive, and financially sustainable.</p><p>Traditional search engines were primarily designed to help users discover information through links to different websites. AI-powered platforms, on the other hand, are increasingly providing direct answers, summaries, and recommendations. This shift is changing how people interact with information and creating challenges for Google’s existing advertising model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/739/1*8gB2Fqe9tAPhykBu5zq9gw.jpeg" /></figure><p>If Google successfully integrates AI into its search engine and develops effective ways to generate revenue from these experiences, it could strengthen its position in the market. However, if AI-powered search generates less advertising revenue than traditional search, Google may struggle to maintain the profitability it has enjoyed for years.</p><p>Google’s future will therefore depend on how effectively it balances innovation, investment, user experience, and profitability.</p><h3><strong>Conclusion</strong></h3><p>Artificial Intelligence is transforming traditional practices across industries, bringing both opportunities and challenges. In areas such as education, customer support, software development, and information discovery, AI has made processes faster, reduced the time required to complete tasks, and opened doors to new possibilities. However, these changes have also disrupted established business models and created uncertainty for companies that rely on traditional services.</p><p>The search engine industry is now undergoing a similar transformation. Google’s traditional search model relies heavily on advertising, while AI-powered platforms are changing how people discover and consume information. Although AI creates opportunities for growth, it also raises questions about Google’s future advertising revenue, increasing expenditure, and long-term profitability. The company’s recent share price decline highlights the financial challenges that come with adapting to such a major technological shift.</p><p>Google’s challenge is not simply to compete with AI companies, but also to adapt to a future in which the very technology transforming its business could disrupt the business model that made it successful.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=300f856eceef" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Vision-Language Models: When AI Moves Beyond Seeing]]></title>
            <link>https://medium.com/@csi_mjcet/vision-language-models-when-ai-moves-beyond-seeing-7c0aa3edd4c7?source=rss-550736826920------2</link>
            <guid isPermaLink="false">https://medium.com/p/7c0aa3edd4c7</guid>
            <category><![CDATA[vision-language-model]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[multimodal-ai]]></category>
            <dc:creator><![CDATA[CSI MJCET]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 14:06:22 GMT</pubDate>
            <atom:updated>2026-10-24T14:06:22.118Z</atom:updated>
            <content:encoded><![CDATA[<p>Author: <a href="https://medium.com/u/de4b0dd9d165">Mohammed Fawaz Malik</a></p><p>The novelty of asking AI what is in a picture has worn off.</p><p>What once felt like a breakthrough is now a fairly ordinary feature of modern AI assistants. We upload documents, analyze screenshots, share our screens, discuss photographs, and interact with models that can process visual information alongside text, audio, and video.</p><p>But beneath this familiarity, something more consequential is happening.</p><p>The question is no longer simply whether artificial intelligence can recognize what it sees. It is whether it can follow what is happening, understand what matters, and make reliable decisions based on that understanding.</p><p>Consider an AI system that can watch a software workflow unfold, identify where an automated process goes wrong, navigate the interface to investigate the problem, and verify whether its intervention worked. Or a robot that can interpret a changing environment, plan a sequence of movements, and adjust those movements when the situation changes.</p><p>These are not merely improvements in image recognition. They represent a shift from visual understanding toward perception-driven intelligence.</p><p>At the centre of this shift lie Vision-Language Models, or VLMs, the systems connecting visual perception with the reasoning and language capabilities of modern AI.</p><p>And their evolution may fundamentally change how we interact with machines.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/939/1*ASCCtiRJEU0x02kdUe6WUQ.png" /></figure><h3><strong>Beyond the Image: The New Frontier of VLMs</strong></h3><p>Traditional computer vision was largely built around specialized tasks. A model might detect pedestrians, classify medical images, recognize handwritten characters, or identify defects on a manufacturing line.</p><p>These systems could achieve remarkable accuracy within their intended domains. However, changing the task often required additional training, new datasets, or an entirely different model.</p><p>Modern VLMs offer a more flexible approach.</p><p>They can combine visual information with natural-language instructions, allowing one model to perform different tasks depending on the context. A chart can become an object of analysis, a document can become a searchable knowledge source, and a screenshot can become the starting point for an interactive workflow.</p><p>The more important development, however, is the growing integration of these capabilities into broader AI systems.</p><p>Google’s Gemini Live, for instance, supports real-time conversations involving camera input and screen sharing. Its <a href="https://blog.google/products-and-platforms/products/gemini/gemini-live-camera-how-to/">2026 guidance on live visual assistance</a> reflects how visual interaction is becoming a practical interface rather than an isolated demonstration.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/494/1*j4ox-T79KgJbWls1ouvdeQ.png" /></figure><p>Meanwhile, advances in computer-use agents allow AI systems to interpret graphical interfaces and interact with digital environments.</p><p>The distinction is important.</p><p>A model that can describe a screen is useful. A system that can understand the screen, carry out a task, and verify the outcome is considerably more powerful.</p><p>This is where VLMs begin intersecting with the broader movement toward agentic AI.</p><h3><strong>Under the Hood: How Visual Intelligence Works</strong></h3><p>Despite their increasingly sophisticated capabilities, many modern VLMs follow an architecture built around three major components.</p><p>The vision encoder processes images and converts them into numerical representations. Modern systems frequently employ Vision Transformers, which divide images into patches and learn visual features across them. Depending on the architecture, these representations may capture everything from textures and object boundaries to higher-level semantic relationships.</p><p>The multimodal connector bridges the vision encoder and the language model. Visual features must be transformed into a representation the language model can interpret. This is often accomplished through a projector or a more sophisticated feature-merging mechanism.</p><p>The language model combines the visual information with the user’s instructions and other available context to produce an answer, plan, or structured output.</p><p>This is a common design pattern, not a universal blueprint. Some models use cross-attention, others process visual information through different tokenization strategies, and newer systems increasingly integrate multiple modalities more deeply into the same architecture.</p><p>The broader evolution is documented in the 2026 survey, <a href="https://www.sciencedirect.com/science/article/pii/S2590005626000627">A Systematic Review of Vision Language Models</a>, which examines architectural approaches, applications, and persistent challenges in building robust multimodal intelligence.</p><p>The important change is that visual information no longer has to remain isolated inside a specialized computer vision pipeline.</p><p>It can become part of the same reasoning process that handles language, retrieves information, uses tools, and plans actions.</p><h3><strong>The New Battleground Is Time</strong></h3><p>An image provides a snapshot of the world. A video introduces sequence, change, and causality.</p><p>That difference complicates the problem considerably.</p><p>A model analyzing a video must determine not only what appears in individual frames, but also how objects move, how events relate to one another, and which moments matter to the question being asked.</p><p>For instance, identifying a person in a video is a relatively straightforward visual task. Determining what caused a machine to malfunction across several minutes of footage is a much harder one.</p><p>The model must preserve relevant context, establish temporal relationships, and avoid confusing correlation with causation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/244/1*lbNynT6UdyhU5gKl3XMqBA.png" /></figure><p>Recent models are increasingly designed for this kind of long-context multimodal processing. Qwen3-VL, introduced in November 2025, supports interleaved text, image, and video inputs within a native 256K-token context window. Its technical report describes improvements to spatial-temporal representations, visual feature integration, and temporal alignment for video.</p><p>These developments matter because the industry is moving beyond one-off visual queries toward the analysis of longer, richer streams of information.</p><p>Think of a manufacturing system that investigates an entire production sequence instead of classifying isolated frames. Or an educational assistant that examines a recorded lecture, identifies the difficult concepts, and connects them to particular moments in the video.</p><p>The opportunity is not simply to process more frames.</p><p>It is to understand what changes, why it matters, and how those changes connect to the larger task.</p><h3><strong>From Understanding to Action</strong></h3><p>Perhaps the most significant development is the transition from visual question answering to visual agency.</p><p>A traditional VLM receives an image and generates a response. An agentic system can use its visual understanding to select actions, interact with tools, and evaluate the results.</p><p>Computer-use agents illustrate this progression particularly well.</p><p>A digital assistant tasked with preparing a report might need to inspect a spreadsheet, identify relevant figures, navigate a presentation tool, construct charts, and save the finished document. Visual perception helps it understand the interface, while planning and tool use help it progress through the workflow.</p><p>This combination is already an active research and product-development area, explored by systems such as <a href="https://openai.com/index/computer-using-agent/">OpenAI’s Computer-Using Agent</a> and Anthropic’s <a href="https://www.anthropic.com/news/3-5-models-and-computer-use">computer-use capabilities</a>.</p><p>The challenge is not merely making the correct prediction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/491/1*hdnzkervx3XIzS8bghFhFA.png" /></figure><p>The system must translate its understanding into the correct interaction, recover from errors, and establish whether the task was actually completed.</p><p>The same principle extends into the physical world.</p><p>Google DeepMind’s <a href="https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/">Gemini Robotics 1.5</a>, announced in September 2025, combines visual and linguistic understanding with action generation for robots. Its accompanying embodied-reasoning model supports spatial understanding, planning, and tool use.</p><p>This is where the distinction between a VLM and a Vision-Language-Action Model, or VLA, becomes important. A VLM can interpret an environment and reason about it. A VLA is designed to connect visual and linguistic inputs directly to physical actions.</p><p>The emerging direction is clear:</p><p>Perception → Reasoning → Planning → Action → Verification</p><p>Each stage introduces its own challenges, and a failure at any point can compromise the entire task.</p><p>Visual intelligence is becoming more valuable not because machines can describe the world with greater fluency, but because their descriptions increasingly influence what they do within it.</p><h3><strong>The Quiet Revolution: Smaller Models, Wider Access</strong></h3><p>There is another development that deserves attention.</p><p>Multimodal intelligence is not exclusively a race toward ever-larger models. Efficiency is becoming an equally important part of the equation.</p><p>Large frontier systems can provide sophisticated reasoning, but running them may require substantial computing resources, create latency, and increase deployment costs.</p><p>Smaller VLMs offer a different set of possibilities.</p><p>Hugging Face’s <a href="https://huggingface.co/blog/smolervlm">SmolVLM family</a> includes models with just 256 million and 500 million parameters, demonstrating how lightweight architectures can support visual understanding within relatively modest resource budgets.</p><p>Such systems are particularly interesting for on-device applications, embedded hardware, privacy-sensitive workflows, and environments with limited connectivity.</p><p>Imagine a quality-inspection system that analyses manufacturing images locally, or a mobile assistant that interprets documents without having to send every image to a remote service.</p><p>Smaller models will not necessarily match frontier systems across every task. Their value lies in making visual intelligence more practical under specific constraints.</p><p>This introduces a different competitive question.</p><p>Instead of asking only which model performs best, developers must also consider which model offers the right balance of accuracy, speed, cost, privacy, and hardware requirements. In many real-world applications, that balance may matter more than a leaderboard position.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/836/1*T-HFp6BuP1JpCwdVZJI_Cw.png" /></figure><p><strong>The Shadow Side: Fluent Does Not Mean Correct</strong></p><p>For all their progress, VLMs still struggle with a fundamental problem: the ability to produce convincing answers without adequately grounding those answers in visual evidence.</p><p>A model might identify the contents of an image correctly but miscount the objects. It might recognize a chart while misinterpreting its axes. It might describe a visual relationship that does not exist or overlook a small detail that changes the meaning of the entire scene.</p><p>These are not merely language-generation errors.</p><p>They are failures of perception, grounding, and reasoning.</p><p>Research has repeatedly demonstrated that strong performance on familiar visual question-answering tasks does not guarantee robust visual understanding. Models can struggle with spatial relationships, detailed counting, fine-grained distinctions, and questions that require combining multiple pieces of visual evidence.</p><p>The problem becomes more serious when those outputs guide automated decisions.</p><p>An incorrect image caption may be an inconvenience. An incorrect interpretation in a medical setting, an industrial inspection system, or a physical robot can have consequences far beyond the conversation itself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/323/1*Cggnr6afrb1guwL2wPMOQA.png" /></figure><p>The challenge is especially important in healthcare. VLMs can help researchers investigate medical imagery, summarize findings, and connect visual observations with relevant information. However, promising benchmark performance is not equivalent to demonstrated clinical reliability. Real-world adoption requires domain-specific validation, careful oversight, and evaluation under realistic conditions.</p><p>A broader <a href="https://www.sciencedirect.com/science/article/pii/S2590005626000627">2026 systematic review</a> similarly identifies hallucination control, robustness, interpretability, and generalization as unresolved challenges.</p><p>The underlying issue is simple.</p><p>A model can generate an answer that sounds certain without having sufficient evidence to justify it.</p><p>And as these systems gain the ability to act, the cost of being confidently wrong increases.</p><h3><strong>When the Image Becomes the Attack</strong></h3><p>There is another complication hiding within multimodal intelligence: security.</p><p>AI systems increasingly process screenshots, photographs, documents, webpages, and other content that may originate from untrusted sources.</p><p>That content can contain instructions directed at the model itself.</p><p>An attacker might embed malicious text within an image, hoping that the AI will interpret it as an instruction rather than as content to be analyzed. This is known as image-based prompt injection.</p><p>A <a href="https://arxiv.org/abs/2603.03637">2026 research paper on image-based prompt injection</a> investigated techniques for concealing adversarial instructions in natural images. Its experiments demonstrated that visually embedded instructions could manipulate model behaviour under the tested conditions.</p><p>The implications extend beyond chatbots.</p><p>Consider an agent that can inspect websites, access business documents, and interact with software. If it mistakes malicious content for an authoritative instruction, its visual capabilities may become an avenue for manipulating its actions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/318/1*oBKyYwumc6ghhG9IMXE0Cg.png" /></figure><p>This is not an argument against multimodal AI. It is a reminder that perception and security must evolve together.</p><p>Systems need to distinguish between information they are supposed to analyse and instructions they are authorized to follow. They also need appropriate permission boundaries, safeguards around sensitive actions, and reliable ways to verify what they are doing.</p><p>Giving an AI access to more of the world requires more than teaching it to interpret that world.</p><p>It requires teaching it which parts of that world it should trust.</p><h3><strong>A New Test for Visual Intelligence</strong></h3><p>As the field advances, measuring progress is becoming increasingly complicated.</p><p>A model that performs well on a collection of visual questions may still struggle with an unfamiliar interface, a long video, or a task requiring several coordinated actions.</p><p>This is why benchmarks are expanding beyond conventional image recognition and short-form question answering.</p><p>MMMU-Pro, for example, introduces more demanding tests of multimodal reasoning by filtering questions that can be answered without the image, expanding answer choices, and including a vision-only evaluation setting. Its <a href="https://arxiv.org/abs/2409.02813">research paper</a> highlights the gap between performance on conventional evaluation and more rigorous tests.</p><p>Computer-use evaluation is evolving as well. The original OSWorld benchmark contains 369 tasks involving real applications and multi-step computer workflows. Its 2026 follow-up, <a href="https://arxiv.org/abs/2606.29537">OSWorld 2.0</a>, focuses on longer, more realistic tasks that expose weaknesses in cross-source reasoning, maintaining context, handling changing environments, and verifying completion.</p><p>These evaluations reflect a broader shift in what counts as progress.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/358/1*iFh8nWAIPd2mFu6VnX1P0A.png" /></figure><p>It is no longer enough for a model to produce a plausible answer to an isolated question. We must ask whether it can sustain performance across an entire task, handle unexpected conditions, communicate uncertainty, and recognize when human intervention is necessary.</p><p>The next generation of evaluation will need to measure not only what a model knows, but also how reliably it applies that knowledge when the environment refuses to cooperate.</p><h3><strong>Beyond the Hype: What Comes Next?</strong></h3><p>The direction of VLM development is becoming increasingly clear.</p><p>Real-time multimodal assistants are bringing visual interaction into everyday workflows. Long-context models are expanding the scope of video analysis. Document-focused systems are changing how information is retrieved from visually complex files. Computer-use agents are connecting perception to digital action, while vision-language-action models are extending the same principles into robotics.</p><p>At the same time, smaller models are broadening the range of possible deployments, and increasingly demanding benchmarks are exposing the distance between impressive demonstrations and dependable systems.</p><p>The next phase will likely be shaped by how effectively these capabilities can be combined.</p><p>Can a model maintain context over a long interaction? Can it identify uncertainty before making an important decision? Can it use tools without losing track of the original objective? Can it distinguish observations from assumptions and verify that its actions achieved the intended result?</p><p>These questions are becoming more consequential than whether a model can describe an image accurately.</p><p>The future of VLMs will depend not just on stronger perception, but on the integration of perception with reasoning, memory, action, and safety.</p><h3><strong>The Machine Has Learned to Look. Has It Learned to Understand?</strong></h3><p>For years, computer vision focused on identifying what appeared in an image.</p><p>Modern VLMs have expanded that ambition. They can connect visual information with language, answer contextual questions, examine long sequences, and support systems that interact with both digital interfaces and the physical world.</p><p>Yet the distance between recognizing something and understanding what it means remains significant.</p><p>A machine can detect an object without understanding its role. It can interpret a scene without recognizing what it has missed. It can generate a convincing explanation without having enough evidence to support it.</p><p>These limitations matter because AI is gradually moving from systems that simply provide information toward systems that participate in decisions and actions.</p><p>The challenge, therefore, is not to create machines that can see everything.</p><p>It is to create machines that can interpret what they see responsibly, distinguish evidence from assumption, and recognize the limits of their own understanding.</p><p>Perhaps that is the real frontier of Vision-Language Models.</p><p>Not teaching AI to see more, but teaching it when seeing is not enough.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7c0aa3edd4c7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[OpenAI’s Smartest Model Yet Is One It Doesn’t Fully Trust]]></title>
            <link>https://medium.com/@csi_mjcet/openais-smartest-model-yet-is-one-it-doesn-t-fully-trust-0a813fcf9c3e?source=rss-550736826920------2</link>
            <guid isPermaLink="false">https://medium.com/p/0a813fcf9c3e</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[gpt-6-astra]]></category>
            <category><![CDATA[claude-fable-5]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <dc:creator><![CDATA[CSI MJCET]]></dc:creator>
            <pubDate>Fri, 04 Sep 2026 11:52:22 GMT</pubDate>
            <atom:updated>2026-09-04T11:52:22.461Z</atom:updated>
            <content:encoded><![CDATA[<p>Author: <a href="https://medium.com/u/e53246dc8219">Syed Unais</a></p><p>Two AI labs, two releases, two days apart.</p><p>On September 1st, Anthropic quietly shipped <strong>Claude Fable 5.1</strong>, an upgrade built less around raw intelligence and more around fixing what made the original Fable expensive and interruption-prone. Two days later, OpenAI answered with <strong>GPT-6 Astra</strong>, calling it “the world’s most intelligent and aligned model.” OpenAI president Greg Brockman went further at the press briefing, saying it’s “not unreasonable to feel that we are now in the AGI era.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eRR5RtdRqnac-qIzgNMahw.jpeg" /></figure><p>That’s a big claim to make in one sentence. The benchmarks back some of it. Not all of it, and that gap is the actual story nobody’s writing.</p><h3>Where Astra Actually Wins</h3><p>Astra saturates FrontierMath Tier 4, a brutal math benchmark, at 97.6%. Fable 5.1 clears the same test at 87.8%. On ExploitBench, a cybersecurity stress test, Astra scores a perfect 100% against Claude Opus 5’s 70%; Anthropic hasn’t published a Fable 5.1 score on this one. It also leads on AutomationBench, a workflow-automation test, 41.4% to Fable 5.1’s 31.4%, more than double what GPT-5.6 Sol managed a generation ago. Computer use got faster too. Astra finishes the same OSWorld tasks in about 40 minutes that took its own predecessor 75, a 47% cut in time. The numbers are real, and they cost real money. Astra runs at $10 per million input tokens, the same list price as Fable 5.1. Fable 5.1’s own cost savings come from cheaper cache reads, not a lower base rate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*EoiHN6uhlf2pSOeTw3H5cA.png" /></figure><h3>Where It Quietly Doesn’t</h3><p>Here’s the part the launch narrative didn’t lead with, even though OpenAI’s own numbers include it. On Humanity’s Last Exam with tools, arguably the hardest general reasoning test that currently exists, Fable 5.1 scores 65.0% to Astra’s 57.2%. On the Artificial Analysis Intelligence Index, a broader composite covering agentic work, coding, and reasoning rather than a single test of raw thought, Fable 5.1 leads again, 65.7 against 61.2. Astra didn’t fail to compete here. It lost.</p><h3>The Coding Fight Nobody’s Talking About</h3><p>Everyone is arguing about which model is smarter. Fewer people are talking about the tool actually sitting in developers’ terminals. Anthropic’s Claude Code just overtook <strong>GitHub Copilot in JetBrains’s 2026 Developer Ecosystem Survey</strong>. Reported usage climbed to 39%, against Copilot’s 21%, among the 15,000 professional developers surveyed. Astra brings real coding improvements of its own, better context retention, fewer wasted iterations. But while both labs were racing benchmark charts, the adoption numbers suggest Claude quietly became the tool developers actually reach for, day to day, task to task.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*zpKyqe4UWc75hL-I4mjkiQ.png" /></figure><h3>The “Aligned” Half of the Claim</h3><p>OpenAI’s title for Astra wasn’t just “<em>most intelligent</em>.” It was “<em>most intelligent and aligned.</em>” In internal testing, Astra never once tried to work around a blocked action; GPT-5.6 Sol did nearly half the time without safeguards in place. Fable 5.1 is chasing the same goal from a different angle. Claude Code users are now seeing roughly 60% fewer safety interruptions than before, and biology-related fallbacks dropped 85%. Neither company thinks a smarter model without restraint is actually a usable one.</p><h3>The Number That Should Worry You More Than Impress You</h3><p>Astra is OpenAI’s first model to cross what it calls the “<em>Critical</em>” cybersecurity threshold under its own Preparedness Framework. During evaluation, run without production safeguards, it discovered and used two previously unknown zero-day vulnerabilities. OpenAI is disclosing both to the affected software maintainers. Its most advanced cyber capabilities, like generating working exploits, stay switched off at launch, even for trusted testers. OpenAI plans to open some of that up gradually through Daybreak, its access-controlled program for authorized security work. Daybreak isn’t simply a badge for “<em>defense-focused organizations</em>.” It spans standard defensive work up through separately vetted, higher-risk tasks. By OpenAI’s own framing, the model just crowned the smartest in the world is also the first one its own creator didn’t fully trust to hand over.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*ZrhnBUxm-CN2k6_HqR0Y-A.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0a813fcf9c3e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[THE SHIFT FROM CLOUD COMPUTING TO EDGE COMPUTING.]]></title>
            <link>https://medium.com/@csi_mjcet/the-shift-from-cloud-computing-to-edge-computing-bd9c49d9e986?source=rss-550736826920------2</link>
            <guid isPermaLink="false">https://medium.com/p/bd9c49d9e986</guid>
            <category><![CDATA[edge-computing]]></category>
            <category><![CDATA[cloud]]></category>
            <category><![CDATA[cloud-computing]]></category>
            <category><![CDATA[cloud-services]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[CSI MJCET]]></dc:creator>
            <pubDate>Mon, 04 May 2026 14:49:03 GMT</pubDate>
            <atom:updated>2026-05-04T14:49:03.358Z</atom:updated>
            <content:encoded><![CDATA[<p>Author: <a href="https://medium.com/u/3661e348b0d9">Faiza Mahveen</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/896/1*6MzY_olQOugSod0FlwN9Pw.png" /></figure><p><strong><em>Is Edge Computing Quietly Replacing the Cloud?</em></strong></p><p>The cloud isn’t going anywhere. But a growing number of companies are routing around it and their reasons say a lot about where computing is headed.</p><p>For most of the last decade, the cloud was the obvious answer. You had data, you sent it to a server farm in Virginia or Oregon, and something came back. The model worked until it didn’t. Today, a quiet recalibration is happening at the infrastructure level, and it’s being driven not by ideology but by physics, regulation, and the stubborn realities of cost.</p><p>Edge computing processing data at or near the point where it’s generated, rather than shuttling it to a centralised data centre is no longer a niche concept. It’s showing up in factory floors, hospital networks, retail stores, and autonomous vehicles. The question isn’t whether edge computing matters. It’s whether it’s becoming the default, and the cloud the fallback. And if you’ve been paying attention to where enterprise infrastructure dollars are quietly flowing, the answer is more interesting than most headlines suggest.</p><h3>The Latency Problem</h3><p>Speed is the biggest reason companies are moving toward edge computing. A self-driving car, industrial robot, or fraud detection system can’t wait for data to travel to the cloud and back. These systems need decisions in milliseconds, and even the fastest internet can’t overcome the physical delay of long-distance data transfer.</p><p>Cloud latency becomes a hard limit for real-time applications. No matter how much you improve servers or networks, you can’t eliminate that delay completely. Edge computing avoids this by processing data locally on the device or nearby systems. This allows instant decisions without depending on external networks. That’s why factories using computer vision detect defects on-site instead of waiting for a cloud response.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/893/1*SOreUshyJbdcUJPFcCyydw.png" /></figure><h4>Privacy and Regulation</h4><p>The legal landscape has changed the calculus, too. GDPR in Europe, HIPAA in healthcare, and a growing patchwork of state-level privacy laws in the U.S. have made “<em>send everything to the cloud</em>” a legally fraught default. Patient data, biometric identifiers, financial records, and behavioral data all carry regulatory weight that makes centralised cloud processing a compliance risk rather than a convenience.</p><ul><li>Strict laws like GDPR and HIPAA limit how data can be stored and processed</li><li>Sensitive data (health, financial, biometric) increases compliance risk in the cloud</li><li>Edge computing keeps raw data local, reducing legal exposure</li><li>Only processed results are sent to the cloud, not full datasets</li><li>Helps meet data residency and privacy requirements</li><li>Widely used in healthcare for secure, local AI processing</li><li>Enables companies to adopt technology without violating regulations</li></ul><p><strong>Hospitals</strong> are already exploiting this architecture. Running diagnostic AI locally means raw imaging data never needs to cross a network boundary. It satisfies data residency requirements without sacrificing the analytical firepower of machine learning. For industries operating under strict compliance frameworks, that’s not a minor benefit it can be the difference between deploying a technology and shelving it.</p><h4>The Cost Equation</h4><p>Then there’s cost. Cloud expenses can grow quickly, especially due to data transfer (egress) and bandwidth charges. The more data a company sends, the more it pays.</p><p><em>For example, a factory generating huge amounts of sensor data can either send everything to the cloud and pay high costs or process most of it locally. Earlier, local processing wasn’t practical because hardware was expensive and limited but that’s no longer the case.</em></p><p>With modern edge devices like NVIDIA Jetson and chips from Hailo and Kneron, local processing has become more affordable and efficient. So the trade-off is simple: spend more upfront on hardware, but reduce ongoing cloud costs. For companies handling large, steady data workloads, edge computing often ends up being the cheaper option in the long run</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/882/1*EADE3nhB-PzVD8w9J9MakQ.png" /></figure><h4>So Is the Cloud Dying?</h4><p><em>It’s not a cloud vs edge competition.</em></p><p>The cloud is still essential for things like training large AI models, handling unpredictable workloads, storing huge amounts of data, and providing global infrastructure. That isn’t changing.</p><p>What’s changing is the mindset. Earlier, cloud was the default choice. Now, companies first consider processing data locally and only use the cloud when it actually makes sense. Edge computing isn’t replacing the cloud it’s just ending the idea that the cloud is always the obvious solution</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bd9c49d9e986" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[They Didn’t Break In, They Logged In: How the Vercel Breach Happened]]></title>
            <link>https://medium.com/@csi_mjcet/they-didnt-break-in-they-logged-in-how-the-vercel-breach-happened-73f07df86fd2?source=rss-550736826920------2</link>
            <guid isPermaLink="false">https://medium.com/p/73f07df86fd2</guid>
            <category><![CDATA[vercel]]></category>
            <category><![CDATA[data-breach]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[hacking]]></category>
            <category><![CDATA[vercel-breach]]></category>
            <dc:creator><![CDATA[CSI MJCET]]></dc:creator>
            <pubDate>Tue, 21 Apr 2026 10:18:37 GMT</pubDate>
            <atom:updated>2026-04-21T10:18:37.625Z</atom:updated>
            <content:encoded><![CDATA[<p>Author: <a href="https://medium.com/u/0f52ace87ae3">Syedmoinuddinhussaini</a></p><p>Most breaches don’t start where you think they do.</p><p>Not with the company that ends up in headlines and definitely not with some dramatic “<em>hack</em>.”<br>This one didn’t start with Vercel either. It started somewhere quieter in a tool, a login, a small decision that probably didn’t feel important at the time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MPFcbC3-OpWCZ6dg.jpg" /></figure><p>The entry point, according to reports, was<strong> Context.ai</strong>, <strong>a third-party tool</strong> used by a Vercel employee. From there, things unfolded in a way that feels less like a breach and more like a chain reaction. Access to a Google Workspace account. Then into internal systems. Then deeper.</p><p>The group linked to the incident, ShinyHunters, didn’t need to force their way in. They moved through what was already accessible just faster and more deliberately, than anyone expected.</p><p>Some early reports suggest that<em> internal credentials</em> and <em>tokens</em> may have been<em> exposed </em>and that data linked to employees could be involved. There are even claims about large datasets being offered for sale. None of this is fully verified in detail yet but the direction of it is enough to raise eyebrows.</p><p>Because Vercel isn’t just another company. It sits underneath a significant portion of the modern web, powering frameworks like Next.js, which sees millions of downloads every week. Even if the confirmed impact is currently limited, the potential reach is hard to ignore.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*av9tNom0svn_wl7t.png" /></figure><p>What’s more interesting, though, is the pace. Vercel’s CEO acknowledged how quickly the attackers moved a kind of speed that feels new. Not necessarily because the techniques are new, but because the execution is. Whether that’s AI-assisted or just highly optimized workflows isn’t entirely clear, but the effect is the same: less time to react, less room for error.</p><p>And maybe that’s the part that lingers.</p><p>This wasn’t about breaking through a wall. It was about finding a side door, something connected, something trusted and walking through it before anyone noticed.</p><p>It’s easy to think of systems as secure because they haven’t failed yet. But incidents like this don’t just test infrastructure. They test assumptions.</p><p>And more often than not, that’s where things actually break.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=73f07df86fd2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Claude Code Leak: Analyzing the Anthropic Code]]></title>
            <link>https://medium.com/@csi_mjcet/the-claude-code-leak-analyzing-the-anthropic-code-96c24c3369dc?source=rss-550736826920------2</link>
            <guid isPermaLink="false">https://medium.com/p/96c24c3369dc</guid>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[source-code-leak]]></category>
            <category><![CDATA[claude-ai]]></category>
            <dc:creator><![CDATA[CSI MJCET]]></dc:creator>
            <pubDate>Wed, 01 Apr 2026 14:42:37 GMT</pubDate>
            <atom:updated>2026-04-01T14:43:36.276Z</atom:updated>
            <content:encoded><![CDATA[<p>Author : <a href="https://medium.com/u/e9342f866bb0">Gazalahafsa</a></p><p>In a world driven by code, a minor mistake just revealed more than expected.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*_AqPX0vSjSKU5EHwqWc5qg.png" /></figure><p>On March 31, 2026, AI company Anthropic found itself in a surprising controversy after unintentionally exposing parts of the source code for its “<em>Claude Code</em>” CLI (command-line interface) tool. The leak happened when a source map file was mistakenly published in an npm package. This blunder quickly caught the attention of developers and security researchers.</p><p>The leaked file was not a complete code dump, but a source map, which links compiled or minified code back to its original, readable version. While useful during development, accidentally including such files in production can expose internal logic that should remain private.</p><p>This effectively made the internal structure of Claude Code easily accessible, allowing observers to see how Anthropic’s command-line AI tool is organized. Although it did not directly reveal proprietary models, it still offered valuable insight into the system’s operation, showing how it processes prompts, manages tasks, and interacts with other components.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/880/1*GzLT7jD6OYidX8k8kArplg.png" /></figure><h3>What caused this mistake?</h3><p>The issue occurred due to a common but serious error in software development. The development source map file was included in a public npm release. The package was published without removing or limiting access to this file. Developers examining the package were able to reconstruct parts of the original source.</p><h4>Why does this matter?</h4><h4>1. Security Concerns</h4><p>Even limited exposure of source code can pose significant risks. It may reveal details about internal APIs and system architecture, suggest possible vulnerabilities, and make it easier for malicious actors to find and exploit weaknesses.</p><h4>2. Competitive Implications</h4><p>Operating in a highly competitive AI environment with companies like OpenAI and Google, Anthropic faces greater risks. Even an accidental leak can give competitors a clearer look at product design and potentially weaken proprietary advantages.</p><h4>3. Trust and Accountability</h4><p>Incidents like this raise concerns about the effectiveness of internal safeguards. They lead to increased scrutiny of security practices, code reviews, and release processes, along with the challenge of balancing transparency and the need to protect sensitive systems.</p><p>Tech media outlets, including NDTV, reported on the leak, emphasizing both the technical nature of the mistake and its broader implications. Some viewed it as a warning about deployment practices, while others found it intriguing to see inside Anthropic’s tooling. Security experts called for stricter CI/CD safeguards.</p><p>With Anthropic’s code exposed, they might take steps to address the issue, which may include:</p><ul><li>Immediate removal of the exposed source code</li><li>Rotation of any exposed credentials</li><li>Strict checks before uploading</li><li>Auditing internal publications</li></ul><p>In the end, this incident serves as a strong reminder that in the world of AI, even small oversights can have significant consequences. In an industry based on code, a single unintended leak can turn even the most sophisticated systems into something surprisingly transparent, like an open book.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=96c24c3369dc" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>