<?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 Pooja Borade on Medium]]></title>
        <description><![CDATA[Stories by Pooja Borade on Medium]]></description>
        <link>https://medium.com/@poojaborade7022?source=rss-59a8239c7722------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*J5Rs_-CMOwAEWaKm.jpg</url>
            <title>Stories by Pooja Borade on Medium</title>
            <link>https://medium.com/@poojaborade7022?source=rss-59a8239c7722------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 23:20:15 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@poojaborade7022/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[The Journey Was the Destination: What Toastmasters Taught Me About Growth, Confidence, and…]]></title>
            <link>https://medium.com/@poojaborade7022/the-journey-was-the-destination-what-toastmasters-taught-me-about-growth-confidence-and-0e7a5d02ee37?source=rss-59a8239c7722------2</link>
            <guid isPermaLink="false">https://medium.com/p/0e7a5d02ee37</guid>
            <category><![CDATA[self-improvement]]></category>
            <category><![CDATA[personal-growth]]></category>
            <category><![CDATA[public-speaking]]></category>
            <category><![CDATA[toastmasters]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Pooja Borade]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 02:43:07 GMT</pubDate>
            <atom:updated>2026-11-19T02:43:07.075Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Journey Was the Destination: What Toastmasters Taught Me About Growth, Confidence, and Leadership</h3><p><em>I joined Toastmasters to learn how to speak. I stayed to discover how to grow.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*DH8qZinQRlRZ6l3N" /></figure><p>Have you ever started something with one expectation, only to discover that it gave you something much more meaningful?</p><p>Perhaps you joined a gym to become physically stronger but ended up developing discipline. Maybe you started learning a new skill for your career but discovered a new passion along the way.</p><p>For me, that experience was Toastmasters.</p><p>When I first joined Toastmasters, my goal was simple: become a better public speaker.</p><p>I wanted to communicate confidently, organize my thoughts, stand in front of an audience without feeling nervous, and deliver speeches that people would remember.</p><p>I thought I was joining a public speaking club.</p><p>I didn’t realize I was beginning a journey of self-discovery.</p><p>Today, I am proud to share that I have completed my <strong>Motivational Strategies Pathway</strong> in Toastmasters International.</p><p>But this milestone means much more to me than completing projects or reaching the final level of a learning path. It represents the person I have become through every speech, evaluation, leadership opportunity, and interaction along the way.</p><p>And I want to share a few lessons this journey has taught me.</p><h3>1. Confidence Comes From Taking Action</h3><p>One of the biggest misconceptions about confidence is that we need to feel confident before taking action.</p><p>We tell ourselves:</p><ul><li>I’ll speak when I’m more prepared.</li><li>I’ll volunteer when I feel ready.</li><li>I’ll take the leadership opportunity when I have enough experience.</li></ul><p>But Toastmasters taught me something different.</p><p><strong>Confidence doesn’t come before you take action. Confidence comes because you take action.</strong></p><p>Your first speech might not be perfect. Your voice might shake. You might forget a few words or lose track of your timing.</p><p>And that’s okay.</p><p>Every time you stand in front of an audience, you take another step outside your comfort zone.</p><p>Over time, those small steps become significant changes.</p><p>I learned that growth doesn’t happen when everything feels comfortable. It happens when we are willing to try something that challenges us.</p><h3>2. Feedback Is a Gift, Even When It Is Difficult to Hear</h3><p>One of the most valuable parts of Toastmasters is its evaluation culture.</p><p>After delivering a speech, you receive feedback from fellow members. They highlight your strengths and identify opportunities for improvement.</p><p>Initially, receiving feedback can feel overwhelming.</p><p>You finish a speech feeling proud, and someone tells you that you could improve your eye contact, gestures, vocal variety, transitions, and conclusion!</p><p>You start wondering whether anything went right.</p><p>But eventually, I understood something important.</p><p>Feedback isn’t criticism of who you are. It’s guidance for who you can become.</p><p>I learned to listen without becoming defensive, appreciate different perspectives, and focus on continuous improvement.</p><p>This lesson extends far beyond public speaking.</p><p>Whether in professional environments, collaborative projects, or personal relationships, the ability to receive feedback and act on it is essential for growth.</p><h3>3. Leadership Is About Helping Others Find Their Voice</h3><p>My Toastmasters journey gradually introduced me to something beyond public speaking: leadership.</p><p>I had the opportunity to serve as Vice President Education at San Jose AI Simplifiers and later as Club President at Santa Clara Entrepreneurs Toastmasters.</p><p>These experiences changed my understanding of leadership.</p><p>When delivering a prepared speech, you have a structure. You know your opening, your key points, and your conclusion.</p><p>Leadership is different.</p><p>It involves people with different personalities, expectations, ideas, and schedules.</p><p>It requires communication, patience, collaboration, and the ability to motivate others.</p><p>I learned that leadership isn’t about being the person who speaks the most or has all the answers.</p><p><strong>Leadership is about creating an environment where others feel encouraged to contribute, grow, and succeed.</strong></p><p>Sometimes leadership means taking responsibility. Sometimes it means stepping back and allowing someone else to lead.</p><p>And sometimes it simply means encouraging someone who doesn’t yet believe in their own abilities.</p><p>The greatest satisfaction comes when you see someone else become more confident because you gave them an opportunity.</p><h3>4. Stop Waiting for Perfection</h3><p>One of the projects I worked on during my journey explored a topic that resonated deeply with me: <em>The Futile Search for Perfection.</em></p><p>It made me reflect on how often we delay opportunities because we are waiting for everything to be perfect.</p><p>The perfect speech.</p><p>The perfect vocabulary.</p><p>The perfect presentation.</p><p>The perfect timing.</p><p>The perfect moment.</p><p>But perfection can become an excuse that prevents us from starting.</p><p>Toastmasters helped me understand that being a work in progress is not a weakness.</p><p>You can be confident and still nervous.</p><p>You can be a leader and still learning.</p><p>You can be experienced and still make mistakes.</p><p>You can celebrate your achievements while continuing to work toward bigger goals.</p><p>I started applying this mindset beyond Toastmasters, too.</p><p>In my academic journey, technical projects, professional development, and interactions with new people, I realized that I didn’t need to know everything before getting started.</p><p>I needed the willingness to learn and the courage to begin.</p><p><strong>Progress is not about being perfect. It’s about being better than you were yesterday.</strong></p><h3>5. Why Human Communication Matters More Than Ever in the Age of AI</h3><p>Being part of San Jose AI Simplifiers makes this lesson particularly interesting.</p><p>We live in a world where artificial intelligence can generate content, analyze information, summarize documents, and even help us prepare speeches.</p><p>AI can probably write a speech faster than I can!</p><p>But there is something technology cannot replace: the human experience behind our words.</p><p>AI can generate a motivational message, but it cannot live your experiences.</p><p>It can suggest a leadership strategy, but it cannot replace the empathy required to understand another person.</p><p>It can generate a presentation, but it cannot replace the courage it takes to stand in front of an audience and share something meaningful.</p><p>As technology continues to evolve, I believe human skills will become even more important.</p><p>Communication, empathy, active listening, leadership, and the ability to inspire others are not simply professional advantages.</p><p>They are essential life skills.</p><p>I believe the future isn’t about choosing between technology and human connection.</p><p>It’s about using technology while continuing to strengthen the qualities that make us human.</p><h3>6. The Biggest Transformation Was Personal</h3><p>So, what has changed since I first joined Toastmasters?</p><p>Am I a perfect speaker? Absolutely not.</p><p>Do I still get nervous? Of course!</p><p>But I am more comfortable expressing my thoughts. I am more willing to take opportunities. I am more open to feedback, and I have developed a deeper appreciation for collaboration and leadership.</p><p>Most importantly, I no longer see public speaking as simply standing in front of people and delivering words.</p><p>I see it as an opportunity.</p><p>An opportunity to share an experience, encourage someone, challenge an idea, make someone laugh, or make someone think differently.</p><p>And perhaps that is the most meaningful transformation.</p><p>I joined Toastmasters to find my voice, but along the way, I discovered the importance of helping others find theirs.</p><h3>7. A Thank You to the People Who Were Part of This Journey</h3><p>No meaningful journey happens entirely on its own.</p><p>Behind every speech, there were fellow Toastmasters who listened, evaluated, encouraged, and challenged me.</p><p>There were mentors who shared their experiences, club members who created opportunities, and leaders who trusted me with responsibilities.</p><p>To everyone who has been part of my Toastmasters journey, thank you.</p><p>Thank you for listening to my speeches, giving me honest feedback, encouraging me when I doubted myself, and creating an environment where learning and growth were always welcomed.</p><p>I am especially grateful to the communities at San Jose AI Simplifiers and Santa Clara Entrepreneurs Toastmasters for the opportunities to learn, contribute, and grow.</p><p>Completing this pathway is my achievement, but the experiences that made it meaningful were shared with so many wonderful people.</p><h3>The Journey Doesn’t End Here</h3><p>Today, I have completed my Motivational Strategies Pathway.</p><p>But I don’t see this as the end of my Toastmasters journey.</p><p>There will be new stages, new challenges, new opportunities, and certainly many more speeches!</p><p>When I look back, I don’t just see completed projects, speeches, or leadership roles.</p><p>I see growth.</p><p>I see the confidence I gradually developed, the lessons I learned from feedback, the friendships I built, and the opportunities I embraced.</p><p>If I could give one piece of advice to someone beginning their Toastmasters journey, it would be this:</p><p>Don’t rush to complete your pathway. Don’t focus so much on the next level that you forget to appreciate the experience you’re having today.</p><p>Enjoy the speeches that challenge you.</p><p>Appreciate the evaluations that make you think.</p><p>Embrace the leadership opportunities that push you outside your comfort zone.</p><p>And most importantly, enjoy the people you meet along the way.</p><p>Because one day, you might look back and realize that the journey you were so eager to complete was transforming you all along.</p><p><strong>I joined Toastmasters to learn how to speak. I stayed to discover how to grow. And I completed this pathway knowing that the greatest achievement wasn’t completing the journey. It was becoming the person I am today because of it.</strong></p><p>Because sometimes, the destination isn’t the certificate, the title, or the applause.</p><p><strong>The journey itself is the transformation.</strong></p><p><em>Here’s to continuous learning, meaningful connections, and becoming a little better every day.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0e7a5d02ee37" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Happens When AI Can Shop for You? Welcome to the Era of Agentic Commerce ]]></title>
            <link>https://medium.com/@poojaborade7022/what-happens-when-ai-can-shop-for-you-welcome-to-the-era-of-agentic-commerce-354b380a2d9b?source=rss-59a8239c7722------2</link>
            <guid isPermaLink="false">https://medium.com/p/354b380a2d9b</guid>
            <category><![CDATA[ecommerce]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Pooja Borade]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 18:45:14 GMT</pubDate>
            <atom:updated>2026-11-15T18:45:14.044Z</atom:updated>
            <content:encoded><![CDATA[<p><em>AI is moving beyond recommending products. It can now help you make purchases, handle payments, and take action on your behalf. But are we ready to let AI spend our money?</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*bnT1Ckgaqjkem49l" /></figure><p>Imagine this.</p><p>You open your AI assistant and type:</p><p>“I need a new laptop for programming, preferably under $1,000. Compare the best options, check their specifications, find the best deal, and order the one that offers the best value.”</p><p>A few years ago, AI would have given you a list of recommendations.</p><p>Today, we’re moving toward something much bigger.</p><p>Your AI assistant can research products, compare prices, help you make a decision, and even initiate the purchasing process.</p><p>Welcome to <strong>Agentic Commerce</strong> — a new chapter in the evolution of artificial intelligence.</p><p>And it might completely change how we shop online.</p><h3>🚀 From Searching for Products to Having AI Buy Them</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/617/0*AKUWgWLdSGptt199" /></figure><p>For years, online shopping has followed a familiar process:</p><ol><li>Search for a product.</li><li>Open multiple websites.</li><li>Compare prices and reviews.</li><li>Add the product to your cart.</li><li>Enter payment information.</li><li>Complete checkout.</li></ol><p>We have become so accustomed to this process that we rarely question it.</p><p>But what if AI could handle most of these steps?</p><p>That’s the idea behind agentic commerce.</p><p>Unlike traditional AI chatbots that primarily answer questions, AI agents can take actions toward a specific goal.</p><p>You give them an objective, provide the necessary permissions, and they work toward completing it.</p><p>Instead of asking AI, “Which headphones should I buy?”</p><p>You could eventually say:</p><p>“Find wireless headphones under $150 with excellent noise cancellation, compare customer reviews, and purchase the best option.”</p><p>The difference might sound small, but technologically, it’s significant.</p><p><strong>AI is moving from helping us make decisions to helping us execute them.</strong></p><h3>💳 Meet ChatGPT Wallet: AI Is Entering the Payment Industry</h3><p>One of the most interesting developments happened on November 13, 2026.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*roX1keiHieHIBW16.png" /></figure><p>OpenAI introduced <strong>ChatGPT Wallet</strong>, a new payment experience designed to let AI agents make purchases on behalf of users.</p><p>This isn’t simply another shopping recommendation feature.</p><p>It’s an attempt to connect AI agents directly with the payment ecosystem.</p><p>Here’s what makes it interesting:</p><p><strong>1. Add your payment card once</strong></p><p>Users can add an eligible payment card to ChatGPT Wallet rather than repeatedly entering their payment details.</p><p><strong>2. Let your AI agent help with shopping</strong></p><p>Your agent, called a dot, can help research products and work toward completing a purchase.</p><p><strong>3. Review before paying</strong></p><p>Before a purchase happens, users can review the selected items, total amount, and payment card.</p><p><strong>4. Approve securely</strong></p><p>Every purchase requires a security check, such as Face ID or fingerprint authentication.</p><p><strong>5. Protect your actual card details</strong></p><p>The system uses payment tokens instead of sharing your actual card details with the agent or merchant.</p><p>OpenAI developed this experience in partnership with major payment networks, including Visa, Mastercard, and American Express.</p><p>There is one important detail: the initial rollout is limited to eligible US ChatGPT Pro subscribers aged 18 and older using eligible US-issued cards through dots.</p><p>This is an early-stage rollout, not a feature available to every ChatGPT user.</p><p>And here’s what makes the future even more interesting.</p><p>OpenAI says it wants to introduce configurable spending limits, purchase categories, and time-based permissions so users can eventually delegate certain purchases to their agents.</p><p>Imagine telling your AI:</p><p>“Whenever my printer ink runs low, find a compatible replacement under $30 and purchase it.”</p><p>That’s a very different shopping experience.</p><h3>🧠 Why Is This Such a Big Deal?</h3><p>Let’s look beyond ChatGPT Wallet.</p><p>Agentic commerce could transform several industries.</p><p><strong>1. E-commerce</strong></p><p>Imagine an AI agent that understands your preferences, compares products across retailers, checks availability, and helps complete your purchase.</p><p>Shopping could become less about browsing endless product listings and more about describing what you need.</p><p><strong>2. Banking and FinTech</strong></p><p>AI agents will need secure ways to interact with payment systems.</p><p>This creates opportunities for payment providers, fraud detection companies, identity verification platforms, and financial technology developers.</p><p><strong>3. Retail and Customer Experience</strong></p><p>Businesses may need to redesign their websites and systems so AI agents can interact with product catalogs, inventory, shopping carts, and checkout processes.</p><p>The next generation of e-commerce websites might need to serve two audiences:</p><ul><li>Human customers.</li><li>AI agents shopping on behalf of human customers.</li></ul><p><strong>4. Personal Productivity</strong></p><p>Think about recurring purchases, travel essentials, office supplies, and household products.</p><p>Instead of repeatedly managing small purchasing decisions, users could eventually delegate routine tasks within predefined limits.</p><p>The real value isn’t just saving money.</p><p>It’s saving time and reducing the number of decisions we need to make every day.</p><h3>🔐 But Can We Really Trust AI With Our Money?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/930/0*h0Bxl-EuggLWRaot.jpg" /></figure><p>This is where things become interesting.</p><p>Giving AI access to information is one thing.</p><p>Giving AI permission to spend money is something entirely different.</p><p>Consider a few questions:</p><ul><li>What happens if an AI agent purchases the wrong product?</li><li>What if it misunderstands your instructions?</li><li>How do we prevent fraudulent transactions?</li><li>Who is responsible when an agent makes an expensive mistake?</li><li>How much personal financial information should an AI system access?</li></ul><p>These aren’t hypothetical concerns that businesses can ignore.</p><p>As AI agents become more capable, security and trust will become just as important as intelligence.</p><p>Tokenized payments, biometric verification, transaction records, spending limits, and user approval are important steps toward safer agentic commerce.</p><p>But technology alone cannot eliminate every risk.</p><p>We still need transparent permissions, clear accountability, and meaningful human control.</p><p><strong>The smartest AI isn’t necessarily the one that can do everything. It’s the one that knows when it needs your approval.</strong></p><h3>🌍 The Bigger Picture: We’re Entering the Age of AI-Powered Transactions</h3><p>Think about how quickly our relationship with technology has evolved.</p><p>A decade ago, we searched Google for information.</p><p>Then we started asking AI chatbots questions.</p><p>Next, AI began helping us write, code, analyze, create, and automate.</p><p>Now, AI agents are beginning to interact with real-world services and financial systems.</p><p>This is a significant shift.</p><p>The internet was originally designed around people navigating websites, clicking buttons, and completing transactions.</p><p>Agentic commerce introduces a different possibility: people communicate their intentions, and software agents handle more of the execution.</p><p>That could change how businesses attract customers, how payment systems work, and how people interact with the digital economy.</p><p>And I think we’re only seeing the beginning.</p><h3>💭 My Take: The Future of Shopping Might Not Involve Shopping at All</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jpOpCSxGlVbhgewS.png" /></figure><p>Here’s something worth thinking about.</p><p>What if, in the future, we rarely visit shopping websites ourselves?</p><p>What if our AI assistant already understands our preferences, knows our budget, compares available options, and brings us only the decisions that genuinely require our attention?</p><p>We might not spend an hour searching for a laptop.</p><p>We might simply describe what we need and review the best recommendation.</p><p>However, I don’t believe convenience should come at the cost of control.</p><p>Our money, preferences, and decisions should remain ours.</p><p>AI should make our lives easier, not make important decisions without our understanding or permission.</p><p>The next big competition in artificial intelligence might not just be about which company builds the smartest model.</p><p>It could be about which company builds the AI agent people trust enough to let it act on their behalf.</p><p>And that is a much bigger challenge than building another chatbot.</p><p><strong>One question to leave you with:</strong></p><p>If your AI assistant could manage your shopping, compare prices, and make purchases within a budget you set, would you trust it enough to let it handle your next online order?</p><p>I’d love to hear your thoughts.</p><p>#ArtificialIntelligence #AIAgents #AgenticCommerce #ChatGPT #FinTech #FutureOfTechnology</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=354b380a2d9b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What If AI Could Think Like a Security Researcher? Exploring OpenAI Codex Security]]></title>
            <link>https://medium.com/@poojaborade7022/what-if-ai-could-think-like-a-security-researcher-exploring-openai-codex-security-b36514a54598?source=rss-59a8239c7722------2</link>
            <guid isPermaLink="false">https://medium.com/p/b36514a54598</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[cloud-security]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[openai-codex]]></category>
            <dc:creator><![CDATA[Pooja Borade]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 17:22:34 GMT</pubDate>
            <atom:updated>2026-10-26T17:22:34.413Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How AI is changing vulnerability detection, threat modeling, and secure software development.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*L6rtstfOZLCMrrss.png" /></figure><p>Imagine discovering that your application has a security vulnerability. You receive an alert, inspect the affected code, investigate whether an attacker could exploit it, reproduce the problem, develop a fix, and test the changes before releasing them.</p><p>Now imagine having an AI agent assist with that entire process.</p><p>This is the direction cybersecurity is moving toward.</p><p>As AI accelerates software development, security teams face a growing challenge: code can be generated faster than humans can thoroughly review it. Traditional security tools remain valuable, but identifying a potential vulnerability is only the beginning. Teams must still determine whether the issue is real, understand its impact, and find a safe way to fix it.</p><p>OpenAI’s <strong>Codex Security</strong> aims to help with this challenge by using AI to investigate vulnerabilities in the context of a software project rather than simply producing a list of suspicious code patterns.</p><p>Introduced as a research preview on March 6, 2026, Codex Security represents an interesting development in AI-assisted application security.</p><p>But what makes it different from conventional security scanners? How does it work? And could AI become an important part of the modern DevSecOps workflow?</p><p>Let’s explore.</p><h3>1. What Is Codex Security?</h3><p>Codex Security is an AI-powered application security tool developed by OpenAI. It is designed to help security and engineering teams identify, validate, and remediate vulnerabilities in software.</p><p>Unlike a tool that only flags potentially dangerous code, Codex Security analyzes a repository to understand how the application works, what it trusts, and where its most important security risks might exist.</p><p>It can then investigate potential vulnerabilities, attempt to validate findings in an isolated environment, and propose targeted fixes for human review.</p><p>The objective is to help security teams spend less time investigating noisy alerts and more time addressing vulnerabilities that genuinely matter.</p><p>According to OpenAI’s official announcement, the product builds on an earlier project called Aardvark.</p><p>You can explore the announcement here:</p><p><a href="https://openai.com/index/codex-security-now-in-research-preview/">Read the official Codex Security announcement</a></p><h3>2. Why Do We Need AI-Powered Security Tools?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NmjJVgVvTujfvnsy" /></figure><p>To understand the value of Codex Security, consider how application security typically works.</p><p>A development team builds an application and uses security tools to identify issues such as vulnerable dependencies, exposed secrets, unsafe input handling, or potentially dangerous code.</p><p>The tools generate findings that security engineers must investigate.</p><p>However, not every finding represents an exploitable vulnerability.</p><p>For example, a scanner might identify a database query that appears vulnerable to SQL injection. But the actual risk depends on how the query is constructed, whether user-controlled input can reach it, what validation occurs elsewhere in the application, and whether the vulnerable code is accessible.</p><p>A security engineer needs to understand this context before deciding what to do.</p><p>This creates several challenges:</p><ul><li><strong>Alert fatigue:</strong> Teams may receive more findings than they can investigate efficiently.</li><li><strong>False positives:</strong> Some alerts turn out not to be exploitable vulnerabilities.</li><li><strong>Limited context:</strong> A code fragment alone may not reveal how a vulnerability affects the overall application.</li><li><strong>Remediation delays:</strong> Even after confirming a vulnerability, developing and testing a safe fix takes time.</li><li><strong>Growing codebases:</strong> More repositories and frequent releases increase the amount of code requiring security review.</li></ul><p>AI agents offer a different approach: investigate the application, reason about how components interact, and help establish whether a suspected vulnerability represents a meaningful risk.</p><p>That is the problem Codex Security is designed to address.</p><h3>3. How Does Codex Security Work?</h3><p>Codex Security follows a workflow that connects understanding, investigation, validation, and remediation.</p><h3>Step 1: Understand the application</h3><p>The system analyzes the repository and creates a project-specific threat model.</p><p>A threat model describes the application’s important assets, entry points, trust boundaries, sensitive data, and potential attack paths.</p><p>For example, imagine a web application with:</p><ul><li>A public API that receives customer requests.</li><li>An authentication service that verifies user identity.</li><li>A backend service that retrieves customer records.</li><li>A database containing sensitive information.</li></ul><p>A useful threat model helps explain how these components interact and where security controls need to be enforced.</p><p>Codex Security’s threat model is editable, allowing teams to correct assumptions and provide additional context about the application’s architecture and deployment.</p><h3>Step 2: Investigate potential vulnerabilities</h3><p>With that context, the system examines the repository for potential security weaknesses.</p><p>Rather than evaluating a line of code in isolation, it can investigate relevant code paths and consider how an issue might affect the application.</p><p>This is particularly useful when a vulnerability depends on the interaction between multiple components.</p><h3>Step 3: Validate the findings</h3><p>A potential vulnerability is not necessarily a confirmed vulnerability.</p><p>Codex Security can attempt to reproduce candidate issues in an isolated environment. Where appropriate, this helps determine whether the suspected weakness is realistic and provides evidence for further investigation.</p><p>The goal is to reduce unnecessary alerts and help security engineers prioritize findings based on evidence and potential impact.</p><h3>Step 4: Propose a targeted fix</h3><p>Once a vulnerability has been investigated, Codex Security can propose a patch intended to address its underlying cause.</p><p>The engineering team can review the proposed changes, examine the supporting evidence, run tests, and decide whether to accept the fix.</p><p><strong>An important distinction:</strong> Codex Security proposes changes for human review. It does not simply make every proposed security change automatically and assume the application is safe.</p><p>This human-review step is essential because a patch that fixes one vulnerability could unintentionally introduce another bug.</p><h3>4. A Practical Example: Investigating a Potential SQL Injection Vulnerability</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/973/0*m5hG2jr7rxuuy4gw" /></figure><p>Let’s consider a hypothetical application that allows users to search for customer records.</p><p>The backend receives a search term and uses it to retrieve matching information from a database.</p><p>Suppose the application constructs a SQL query by directly combining the search term with a query string.</p><p>This could create a SQL injection risk if an attacker can control the input and influence the database query.</p><p>Here is a simplified example of the vulnerable pattern:</p><pre>query = &quot;SELECT * FROM customers WHERE name = &#39;&quot; + name + &quot;&#39;&quot;<br>cursor.execute(query)</pre><p>The problem is that user-controlled data is inserted directly into the SQL statement.</p><p>A safer approach is to use parameterized queries:</p><pre>query = &quot;SELECT * FROM customers WHERE name = ?&quot;<br>cursor.execute(query, (name,))</pre><p>The exact placeholder syntax depends on the database driver.</p><p>How could an AI security agent help investigate this scenario?</p><p><strong>First, it could inspect the code path.</strong> Where does the name value originate? Is it received from a public API? Does it pass through other functions before reaching the database?</p><p><strong>Second, it could evaluate the potential impact.</strong> If untrusted input reaches a database query without safe parameterization, what information or functionality might be exposed?</p><p><strong>Third, it could attempt validation.</strong> In an appropriate isolated test environment, the agent could investigate whether the suspected vulnerability can be reproduced.</p><p><strong>Finally, it could propose remediation.</strong> A targeted change might use parameterized queries and include relevant tests to ensure the behavior remains correct.</p><p>This example illustrates the type of reasoning an AI security workflow can support. It is not a claim that Codex Security has been run against this specific example.</p><p>The important lesson is that finding a suspicious pattern, establishing its actual risk, and fixing its root cause are three different tasks.</p><h3>5. How Could Codex Security Fit Into DevSecOps?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7F5VHSFqzj3jaxct" /></figure><p>DevSecOps brings security practices into the software development and delivery lifecycle instead of treating security as a final checkpoint before deployment.</p><p>AI-assisted security tools could strengthen this approach by helping teams investigate vulnerabilities earlier and respond to findings more efficiently.</p><p>Consider a simplified workflow:</p><ol><li>A developer commits code to a GitHub repository.</li><li>The development team uses its existing automated tests and security checks.</li><li>Codex Security can analyze a repository or relevant commits.</li><li>Potential vulnerabilities are investigated and, where possible, validated.</li><li>The system provides evidence and remediation guidance.</li><li>Engineers review the findings and proposed fixes.</li><li>Approved changes go through the normal testing and pull request process.</li><li>The team verifies that the fix works and does not introduce regressions.</li></ol><p>Codex Security Cloud also supports ongoing repository monitoring, including scans of new commits.</p><p>This creates an opportunity to incorporate AI-assisted security analysis into an existing development workflow rather than waiting for a separate security review at the end.</p><p>However, an organization should still maintain its existing security controls. AI-assisted analysis complements security testing, dependency scanning, secret detection, access controls, and code review rather than replacing every one of them.</p><p>For more information, see <a href="https://openai.com/business/solutions/cybersecurity/">OpenAI’s cybersecurity overview</a>.</p><h3>6. Codex Security vs. Traditional Security Scanners</h3><p>It is useful to understand the difference between traditional application security testing and AI-assisted investigation.</p><p>Traditional tools often excel at detecting known patterns, checking dependencies against vulnerability databases, enforcing rules, or identifying specific classes of security problems.</p><p>AI-assisted tools bring contextual reasoning and investigation to the process.</p><p>CapabilityTraditional security scannersAI-assisted security investigationPattern detectionOften based on rules and signaturesCan reason about broader code contextDependency vulnerabilitiesSpecialized tools are highly usefulMay complement dependency analysisApplication contextVaries by toolCan build a project-specific threat modelValidationDepends on the tool and testing configurationCan investigate and attempt to reproduce findingsRemediationMay provide rules or recommended fixesCan propose targeted patchesHuman reviewImportantStill essential</p><p>These categories are not mutually exclusive. Modern security platforms can combine static analysis, dependency intelligence, dynamic testing, and AI-driven investigation.</p><p>Codex Security’s distinctive emphasis is on using repository context and AI reasoning to investigate potential vulnerabilities and help teams move toward evidence-backed remediation.</p><p>It should not be interpreted as proof that AI will identify every vulnerability or outperform every specialized security scanner.</p><h3>7. What Are the Limitations?</h3><p>Despite its potential, AI-powered application security comes with important limitations.</p><h3>AI can still miss vulnerabilities</h3><p>No security tool can guarantee that an application is free from vulnerabilities. An AI agent may overlook an issue, misunderstand a code path, or fail to reproduce a weakness.</p><h3>Validation is not a guarantee of safety</h3><p>A vulnerability that cannot be reproduced in one environment might still be exploitable under different deployment conditions. Test environments do not always reflect production systems perfectly.</p><h3>Proposed fixes need review</h3><p>A generated patch may solve the immediate problem but introduce a regression, change application behavior, or fail to address a deeper architectural weakness.</p><h3>Threat models can become outdated</h3><p>Applications change as teams add features, modify permissions, integrate new services, and migrate infrastructure. The threat model needs to reflect those changes to remain useful.</p><h3>Access and data handling require attention</h3><p>Connecting a security tool to a source code repository involves permissions and potentially sensitive information. Organizations should review access scopes, workspace policies, data-handling requirements, and scan costs before adopting a new workflow.</p><p>For these reasons, AI-generated findings and fixes should be treated as engineering recommendations that require verification.</p><h3>8. Who Can Benefit From Codex Security?</h3><p>Codex Security is particularly relevant to teams that maintain software repositories and need to improve their application security processes.</p><p><strong>Application security engineers</strong> can use AI-assisted investigation to help prioritize findings and examine potential attack paths.</p><p><strong>Software developers</strong> can receive more contextual security findings and investigate proposed fixes as part of their normal development workflow.</p><p><strong>DevSecOps engineers</strong> can explore ways to integrate security checks into repository monitoring and development processes.</p><p><strong>Cloud engineers</strong> can benefit when securing the application code that supports cloud-hosted services, APIs, and data platforms.</p><p><strong>Engineering teams</strong> can use evidence-backed findings to support conversations between developers and security specialists.</p><p>For students and early-career professionals, learning how these tools work can also provide a practical way to understand threat modeling, vulnerability validation, secure coding, and the software remediation lifecycle.</p><h3>9. How Can You Explore It Yourself?</h3><p>If you want to understand AI-assisted security beyond reading about it, consider a small, controlled experiment.</p><p>Start with a GitHub repository that you own or have explicit permission to assess. A deliberately vulnerable sample application or a non-production project is a good starting point.</p><p>A suggested learning workflow:</p><ol><li>Choose a small application with a few API endpoints and database operations.</li><li>Document its architecture, entry points, sensitive data, and trust boundaries.</li><li>Review its code manually and identify potential security concerns.</li><li>Check whether you have access to Codex Security and connect the appropriate repository.</li><li>Run a repository scan or a commit scan, following the product’s setup instructions.</li><li>Compare the findings with your own review and other appropriate security checks.</li><li>Examine the evidence, validation details, and proposed remediation.</li><li>Review any proposed patch, run relevant tests, and verify the fix before accepting it.</li></ol><p>You can find setup guidance in the <a href="https://help.openai.com/en/articles/20001107-codex-security">official Codex Security documentation</a>.</p><p>Codex Security is a research-preview product, and access depends on eligibility and workspace configuration. Review the current documentation before planning a hands-on test.</p><p>Avoid scanning repositories or systems without authorization, and do not expose production credentials or sensitive customer information during experimentation.</p><h3>10. The Bigger Picture: From Finding Vulnerabilities to Fixing Them</h3><p>One of the most interesting developments in AI-powered cybersecurity is the shift from detection alone toward a more complete investigation and remediation workflow.</p><p>For years, security teams have invested in tools that identify weaknesses. But an alert is only the start of the process.</p><p>Someone must understand the application, establish whether the issue is real, determine its severity, develop an appropriate fix, and verify that the fix works.</p><p>AI agents could help reduce the effort required across these stages.</p><p>Codex Security is an example of this direction: using AI to understand software context, investigate potential vulnerabilities, collect supporting evidence, and propose fixes that developers can review.</p><p>The long-term opportunity is not necessarily to remove humans from application security. It is to help security professionals spend more time on complex decisions and less time on repetitive investigation.</p><p>As software development continues to accelerate, this kind of assistance could become increasingly valuable.</p><h3>Final Thoughts</h3><p>AI-powered coding tools are making it easier to generate software. That creates an equally important challenge: ensuring that the software we build is secure.</p><p>OpenAI Codex Security offers an interesting approach by combining code understanding, threat modeling, vulnerability investigation, validation, and remediation assistance.</p><p>It has potential, but its effectiveness depends on the quality of its analysis, the context provided by the team, the environment used for validation, and the engineering review process around its recommendations.</p><p>For developers and DevSecOps professionals, the most valuable skill will not simply be knowing how to ask AI to write code. It will be knowing how to evaluate that code, understand its risks, and verify that the resulting system behaves securely.</p><p><strong>The future of application security may not be about choosing between human expertise and AI. It may be about combining both to build safer software at the pace modern development demands.</strong></p><p><em>What do you think? Would you trust an AI agent to investigate vulnerabilities in your codebase, or would you prefer to use it only as a second opinion? Share your thoughts in the comments.</em></p><p><strong>Further reading</strong></p><ul><li><a href="https://openai.com/index/codex-security-now-in-research-preview/">Codex Security: Now in Research Preview — OpenAI</a></li><li><a href="https://help.openai.com/en/articles/20001107-codex-security">Codex Security — OpenAI Help Center</a></li><li><a href="https://openai.com/business/solutions/cybersecurity/">AI for Cybersecurity Teams — OpenAI</a></li></ul><p>#ArtificialIntelligence #Cybersecurity #DevSecOps #OpenAI #CloudSecurity #SoftwareEngineering #AITools</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b36514a54598" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Robot That Learns Like You Scroll: Inside Skild AI’s Wild Bet on a Universal Robot Brain]]></title>
            <link>https://medium.com/@poojaborade7022/the-robot-that-learns-like-you-scroll-inside-skild-ais-wild-bet-on-a-universal-robot-brain-95c10fd7c06f?source=rss-59a8239c7722------2</link>
            <guid isPermaLink="false">https://medium.com/p/95c10fd7c06f</guid>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[skild-ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Pooja Borade]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 18:19:49 GMT</pubDate>
            <atom:updated>2026-09-23T18:19:49.211Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Imagine showing a robot a video of someone potting a plant, then watching it do the same thing eleven minutes later.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*sgTwAKwwn2KEKVaG" /></figure><p>No weeks of programming. No engineers driving the robot around by remote control for hours. Just: <em>watch this, now you do it.</em></p><p>That’s the pitch behind <strong>Skild AI</strong>, a Pittsburgh startup that has quietly become one of the most talked-about companies in robotics. Its newest model, S1, might be the closest thing yet to a “ChatGPT moment” for robots.</p><p>Let’s dig in.</p><h3>The Problem Nobody Talks About: Robots Are Weirdly Dumb</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jjpwc9F-WeqiuR-x.jpg" /></figure><p>We’ve all seen the viral demos: the backflipping robot, the dancing robot. But the dirty secret of robotics is that most robots are <em>brilliant at exactly one thing</em> and helpless at everything else.</p><p>Teaching a robot a new task usually means hours of teleoperation, then a task-specific training run. Every new job restarts the cycle. That’s why robots have stayed in tightly controlled factories instead of your kitchen.</p><p>Skild’s answer is to skip the hardware race entirely and build the <strong>brain</strong>.</p><h3>One Brain, Any Body</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/738/0*fpIvs6VPPhiwPUL1" /></figure><p>Here’s what makes Skild unusual. While many robotics companies build their own machines, Skild isn’t building proprietary hardware. It’s developing a robot-agnostic foundation model that can be customized for various types of robots and use cases.</p><p>They call it the <strong>Skild Brain</strong>, and it’s <em>omni-bodied</em>. In the company’s words, it can control any robot without prior knowledge of their exact body form, including quadrupeds, humanoids, tabletop arms, and mobile manipulators.</p><p>Think of it like an operating system. Windows doesn’t care which laptop brand you bought. Skild wants its brain to work the same way, whether the body has four legs or two arms.</p><h3>“But Where Do You Get the Data?”</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QbOgnqBzENOb8zQw.jpg" /></figure><p>Language models had the internet. Robots have… nothing comparable. There is no giant internet of robot movements.</p><p>Skild’s workaround is clever. The company pre-trains its model on two scalable sources: human videos from the internet, and practice inside physics-based simulations. Basically, the robot learns by watching us and by playing millions of rounds in a virtual gym. The simulation work runs on NVIDIA tools, including Isaac Lab for physics simulation and Cosmos for data augmentation and generation.</p><h3>Enter S1: The “Show, Don’t Tell” Robot</h3><p>In late August 2026, Skild unveiled <strong>S1</strong>, and this is where things get exciting.</p><p>S1 uses <em>in-context learning</em>, the same trick that lets you paste examples into a chatbot prompt and get it to follow along. Except here, the prompt is a <strong>video</strong>. CEO Deepak Pathak explained it simply: you add a video of a human doing something to the model’s context, and it can follow it on the robot.</p><p>The demos aren’t toy tasks either. S1 was shown performing previously unseen tasks such as plant potting, pancake cooking, pour-over coffee making and kit assembly, involving dozens of manipulation steps and running up to 10 minutes from a single demonstration. In the plant-potting test, only 11 minutes passed between recording the human demo and the robot starting the task on its own.</p><p>Pathak stressed that these aren’t three-second, four-second tasks, not those tiny, simple tasks. They’re long-horizon jobs where one small mistake can ruin everything.</p><h3>The Money Tells Its Own Story</h3><p>Investors clearly believe. Skild came out of stealth in 2024 with a $300 million Series A that valued the company at $1.5 billion. By January 2026, it had raised close to $1.4 billion in a round led by SoftBank, with NVIDIA’s venture arm and Jeff Bezos participating. The valuation reportedly topped $14 billion, and total funding is now nearly $1.7 billion.</p><p>Some commentators are already calling this a “GPT-3 moment” for physical AI, and one newsletter claims Skild has passed $100 million in revenue this year. That figure comes from a commentator rather than the company’s own filings, so treat it with a pinch of salt.</p><h3>Let’s Stay Curious (and a Little Skeptical)</h3><p>Here’s the honest part. Skild’s own announcement is a company claim, and independent observers note that further technical evaluation will be needed to assess generalisation and production readiness. A polished demo of pancake-flipping isn’t the same as a robot working reliably in a messy, unpredictable home at 2 a.m.</p><p>Big questions remain:</p><ul><li><strong>Reliability:</strong> Can it succeed 99.9% of the time, not just in demo conditions?</li><li><strong>Safety:</strong> What happens when a robot follows a video <em>slightly wrong</em> around people?</li><li><strong>Economics:</strong> Will cheap hardware plus a smart brain beat purpose-built machines?</li></ul><h3>Why This Matters</h3><p>If Skild’s approach holds up, the implication is huge: <strong>the value in robotics shifts from the body to the brain.</strong> Robot makers wouldn’t need to reinvent intelligence for every machine. They’d plug in a brain, show it a video, and go.</p><p>The company says its long-term goal is robots in consumer homes, with enterprise tasks as the first application.</p><p>So the next time you watch a cooking video, remember: someday a robot might be watching it too, taking notes.</p><p><em>What task would you teach a robot first? Drop it in the comments.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=95c10fd7c06f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Is No Longer Just Chatting: Meet Meta Muse, the Agent That Can Actually Do Things]]></title>
            <link>https://medium.com/@poojaborade7022/ai-is-no-longer-just-chatting-meet-meta-muse-the-agent-that-can-actually-do-things-ff0127344eda?source=rss-59a8239c7722------2</link>
            <guid isPermaLink="false">https://medium.com/p/ff0127344eda</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[meta]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[meta-muse]]></category>
            <dc:creator><![CDATA[Pooja Borade]]></dc:creator>
            <pubDate>Wed, 09 Sep 2026 14:52:14 GMT</pubDate>
            <atom:updated>2026-09-09T14:52:14.850Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qUE0l0MbFFDHuuqM" /></figure><p>For the last few years, we have been teaching AI how to <strong>answer our questions</strong>.</p><p>Now, we’re entering a phase where we’re teaching AI how to <strong>do the work</strong>.</p><p>On September 8, 2026, Meta introduced <strong>Muse</strong>, its new personal AI agent. And unlike the AI assistants we’ve become accustomed to, Muse isn’t designed merely to give you a clever response and wait for your next instruction.</p><p>You can give it a goal.</p><p>Then it can figure out the steps.</p><p>It can open a browser, fill out forms, send emails, book travel, make purchases, and continue working even after you close the app.</p><p>That sounds like a small upgrade.</p><p>It isn’t.</p><p>It represents a much bigger change in how we may interact with computers.</p><h3>We’ve Been Asking AI the Wrong Question</h3><p>Think about how you normally use ChatGPT or another AI chatbot.</p><p>You type:</p><blockquote><em>“Find me a good flight to New York.”</em></blockquote><p>The AI gives you suggestions.</p><p>You then open a travel website.</p><p>You compare prices.</p><p>You choose a flight.</p><p>You enter your information.</p><p>You pay.</p><p>The AI helped you <strong>think</strong>.</p><p>But you still had to do the work.</p><p>Now imagine saying:</p><blockquote><em>“Find me the best flight to New York for next weekend. Keep it under $400, choose a reasonable departure time, and let me know when you find one.”</em></blockquote><p>And then walking away.</p><p>That’s the fundamental difference between a chatbot and an AI agent.</p><p><strong>A chatbot responds.</strong></p><p><strong>An agent acts.</strong></p><p>And Muse is Meta’s attempt to bring that idea into everyday consumer life.</p><h3>So, What Exactly Is Meta Muse?</h3><p>Meta describes Muse as a personal AI agent built to help people accomplish goals rather than simply answer questions.</p><p>It can connect to services you already use — including email, calendars, shopping, payments, health and fitness services, smart-home systems, and other apps.</p><p>You can access it through a dedicated Muse app, the web, or WhatsApp, with a rollout initially focused on the United States. Meta also says Muse is coming to its AI glasses.</p><p>But the most interesting part isn’t where you access Muse.</p><p>It’s <strong>what happens after you give it a task.</strong></p><p>Suppose you’re planning a dinner party.</p><p>A traditional AI assistant might give you:</p><ul><li>A menu</li><li>A shopping list</li><li>Recipes</li><li>Suggestions for invitations</li></ul><p>Muse is designed to take things further.</p><p>It could remember a recipe you saved, create a grocery list, consider your guests’ dietary restrictions, help organize the dinner, and potentially handle parts of the process across your connected services.</p><p>The AI isn’t simply generating information anymore.</p><p>It’s becoming an <strong>execution layer between you and the internet</strong>.</p><p>And that’s a very different idea.</p><h3>From Chatbots → Copilots → Agents</h3><p>There’s a useful way to understand where AI is heading.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/0*zET9EYz9FyU1YAi_" /></figure><h3>Generation 1: Chatbots</h3><p>You ask.</p><p>AI answers.</p><p><strong>Human does the work.</strong></p><h3>Generation 2: Copilots</h3><p>You work alongside AI.</p><p>AI writes the email.</p><p>AI summarizes the document.</p><p>AI generates the code.</p><p><strong>Human + AI do the work together.</strong></p><h3>Generation 3: Agents</h3><p>You describe the outcome.</p><p>AI figures out the steps.</p><p>AI interacts with software.</p><p>AI executes tasks.</p><p><strong>Human defines the goal. AI handles more of the execution.</strong></p><p>Muse is an example of this third category.</p><p>And that shift may be more important than another improvement in chatbot intelligence.</p><p>Because productivity isn’t always limited by our ability to generate information.</p><p>Sometimes it’s limited by the number of tiny tasks required to actually accomplish something.</p><h3>The Real Power of Muse Is Not One Task</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oxM_r9NieEN_ICKw" /></figure><p>Sending an email isn’t revolutionary.</p><p>Neither is booking a flight.</p><p>Neither is filling out a form.</p><p>The interesting part is combining all of them.</p><p>Imagine telling an AI:</p><blockquote><em>“I’m moving to Chicago next month. Help me organize everything.”</em></blockquote><p>That’s no longer a simple question.</p><p>It’s a project.</p><p>The agent might need to:</p><ol><li>Create a checklist.</li><li>Research moving services.</li><li>Compare prices.</li><li>Organize dates.</li><li>Manage appointments.</li><li>Update your calendar.</li><li>Find an apartment.</li><li>Track documents.</li><li>Remind you about deadlines.</li><li>Continue working as circumstances change.</li></ol><p>This is where <strong>agentic AI</strong> becomes fundamentally different from generative AI.</p><p>Generative AI creates.</p><p>Agentic AI <strong>coordinates and executes</strong>.</p><p>Meta says Muse is designed to take on longer-term goals, create personalized plans, coordinate resources, and continue working in the background when necessary.</p><p>That is much closer to having a digital assistant than having a chatbot.</p><h3>And Here’s the Slightly Scary Part</h3><p>The more useful an AI agent becomes, the more access it needs.</p><p>That’s the paradox.</p><p>If you want an AI to recommend a restaurant, it doesn’t need much access.</p><p>If you want it to <strong>make the reservation</strong>, it needs access to your calendar.</p><p>If you want it to <strong>buy the ingredients</strong>, it needs access to shopping and payment systems.</p><p>If you want it to <strong>manage your life</strong>, eventually it needs access to a surprisingly large part of your digital life.</p><p>And that’s where things get complicated.</p><p>Meta knows this.</p><p>The company has built Muse around what it calls a <strong>Muse Secure VM</strong> — a dedicated virtual machine where the agent and user data are isolated. A separate system called <strong>Sentinel</strong> monitors interactions with the outside world and can require user approval for sensitive actions.</p><p>Meta also says users decide which apps Muse can access and what permissions it receives.</p><p>For example, you can decide whether Muse can merely read your email or also send messages on your behalf.</p><p>And for purchases, Meta says Muse can use Stripe’s Link system, including one-time-use card information designed to keep a user’s actual card details hidden.</p><p>These protections matter.</p><p>Because giving an AI permission to <strong>write an email</strong> is one thing.</p><p>Giving it permission to <strong>send the email</strong> is another.</p><p>Giving it permission to <strong>spend your money</strong> is something else entirely.</p><h3>The Biggest Challenge Isn’t Intelligence</h3><p>It’s trust.</p><p>We’ve spent years asking:</p><blockquote><em>“How smart is AI?”</em></blockquote><p>The next question may be:</p><blockquote><strong><em>“How much should we let it do?”</em></strong></blockquote><p>Imagine an AI that is 99% accurate.</p><p>That sounds incredible.</p><p>But what happens when the 1% mistake is:</p><ul><li>An email sent to the wrong person</li><li>A purchase you didn’t intend to make</li><li>A reservation canceled accidentally</li><li>A confidential document exposed</li><li>A misleading website successfully tricking the agent</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cN_50Q45EhUzG_YN" /></figure><p>AI agents face a particularly difficult problem called <strong>prompt injection</strong>, where malicious instructions hidden inside webpages, documents, emails, or other content can potentially manipulate an AI system.</p><p>Meta says it designed Muse assuming the agent could be attacked and built multiple layers of protection around it. The company has also opened a public bug bounty program for Muse, with rewards of up to $300,000 for qualifying security reports.</p><p>That tells us something important.</p><p>The future of AI isn’t only about making models smarter.</p><p>It’s about making <strong>autonomous systems reliable enough to trust</strong>.</p><h3>What Makes Muse Different From Just Another AI App?</h3><p>The answer is <strong>agency</strong>.</p><p>A typical AI conversation looks like this:</p><p><strong>You → Prompt → AI → Answer → You</strong></p><p>An agent looks more like:</p><p><strong>You → Goal → AI → Plan → Tools → Actions → Results → Feedback</strong></p><p>That’s a fundamental change.</p><p>Instead of interacting with AI every few minutes, you might eventually give an agent a goal and check in only when something important happens.</p><p>The computer starts working for you.</p><p>Not the other way around.</p><h3>Imagine What This Could Mean for Work</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/0*fP311UisHI1ApL3E" /></figure><p>This is where things get especially interesting.</p><p>Think about the average office job.</p><p>A huge amount of work isn’t really “thinking.”</p><p>It’s coordination.</p><p>Checking emails.</p><p>Updating spreadsheets.</p><p>Scheduling meetings.</p><p>Following up with people.</p><p>Filling forms.</p><p>Searching systems.</p><p>Creating reports.</p><p>Copying information between applications.</p><p>Updating CRM records.</p><p>Tracking deadlines.</p><p>Sending reminders.</p><p>Generating status updates.</p><p>None of these tasks individually look revolutionary.</p><p>But collectively, they consume enormous amounts of time.</p><p>AI agents could potentially take over significant portions of this coordination layer.</p><p>And that could change the definition of productivity.</p><p>The most valuable employee may not be the person who can manually complete 100 tasks.</p><p>It may increasingly be the person who can tell an AI agent:</p><blockquote><em>“Here is the outcome I need. Here’s what you’re allowed to do. Here’s what requires my approval.”</em></blockquote><p>That is a very different skill.</p><h3>The New AI Skill: Delegation</h3><p>For years, we’ve been learning <strong>prompt engineering</strong>.</p><p>Write better prompts.</p><p>Give better context.</p><p>Structure your instructions.</p><p>But agentic AI could shift the skill again.</p><p>The next skill may be <strong>AI delegation</strong>.</p><p>Knowing:</p><ul><li>What should I delegate?</li><li>What should I keep for myself?</li><li>What permissions should I give?</li><li>What decisions require human approval?</li><li>How do I verify the result?</li><li>What happens if the agent makes a mistake?</li></ul><p>In other words, we may move from:</p><p><strong>“How do I ask AI a better question?”</strong></p><p>to:</p><p><strong>“How do I manage an AI that is actually doing work for me?”</strong></p><p>That’s a much bigger change.</p><h3>But Let’s Not Get Carried Away</h3><p>Muse is impressive.</p><p>But it isn’t magic.</p><p>Even Meta acknowledges that agents can make mistakes, and the company’s own safety research describes the challenge of allowing an agent to operate unattended while interacting with real-world systems.</p><p>And there is another problem:</p><p><strong>Context isn’t the same as judgment.</strong></p><p>An AI might understand your instructions perfectly and still make a decision you wouldn’t have made.</p><p>It might find the cheapest flight — but not realize you would rather pay $50 more to avoid a six-hour layover.</p><p>It might create the “perfect” dinner plan — but underestimate how difficult the recipes are for a beginner.</p><p>It might technically accomplish the task while completely missing the human preference behind it.</p><p>That distinction will become increasingly important.</p><h3>The Bigger Picture: Your AI Could Become Your Digital Proxy</h3><p>This is the part that fascinates me most.</p><p>Today, most software waits for you.</p><p>You open Gmail.</p><p>You search.</p><p>You click.</p><p>You type.</p><p>You submit.</p><p>You open a shopping website.</p><p>You search.</p><p>You compare.</p><p>You purchase.</p><p>The interface is designed around <strong>you operating the computer</strong>.</p><p>Agentic AI flips that relationship.</p><p>You tell the AI what you want.</p><p>The AI operates the computer.</p><p>That means the traditional app interface may eventually become less important.</p><p>Instead of learning how to use ten different apps, you might simply tell your AI what you want — and let the AI figure out which apps and services are necessary.</p><p>This is one reason the rise of AI agents could be much bigger than another chatbot upgrade.</p><p><strong>The interface to the internet itself could be changing.</strong></p><h3>So, Is the Chatbot Era Over?</h3><p>Not exactly.</p><p>Chatbots aren’t going anywhere.</p><p>There will always be situations where you simply want an answer, an explanation, a brainstorming partner, or someone to help you think.</p><p>But we’re moving toward something bigger.</p><p>The evolution looks something like this:</p><p><strong>Search:</strong><br>“Find information for me.”</p><p>↓</p><p><strong>Chatbots:</strong><br>“Explain this to me.”</p><p>↓</p><p><strong>Copilots:</strong><br>“Help me do this.”</p><p>↓</p><p><strong>Agents:</strong><br>“Do this for me.”</p><p>And that last step changes the relationship completely.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RpwemcIHUsvqL_DO" /></figure><h3>The Real Question Isn’t “Can AI Do It?”</h3><p>We’re reaching a point where the answer increasingly becomes:</p><p><strong>Yes.</strong></p><p>The more interesting question is:</p><blockquote><strong><em>Should it?</em></strong></blockquote><p>Should an AI send your emails?</p><p>Should it negotiate your bills?</p><p>Should it buy things for you?</p><p>Should it manage your calendar?</p><p>Should it access your health information?</p><p>Should it make decisions without asking?</p><p>And eventually:</p><p><strong>Should you trust it enough to let it manage parts of your life while you’re not even looking?</strong></p><p>Those aren’t purely technical questions.</p><p>They’re human questions.</p><p>And that’s why Meta Muse is interesting.</p><p>Not because it is the first AI that can send an email.</p><p>But because it represents a larger transition:</p><p><strong>AI is moving from being something we talk to into something we delegate to.</strong></p><p>The chatbot era taught AI how to communicate with us.</p><p>The agent era is teaching AI how to <strong>act on our behalf</strong>.</p><p>And if that transition succeeds, the next generation of computing may feel less like opening apps…</p><p>…and more like having a digital assistant quietly working in the background.</p><p>The question is no longer whether AI can do things.</p><p><strong>It’s how much of our lives we’re willing to let it do.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ff0127344eda" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Shadow AI: The Security Risk Hiding Inside Modern DevSecOps Pipelines]]></title>
            <link>https://medium.com/@poojaborade7022/shadow-ai-the-security-risk-hiding-inside-modern-devsecops-pipelines-dd53226c7920?source=rss-59a8239c7722------2</link>
            <guid isPermaLink="false">https://medium.com/p/dd53226c7920</guid>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[cloud-security]]></category>
            <category><![CDATA[shadow-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Pooja Borade]]></dc:creator>
            <pubDate>Mon, 17 Aug 2026 15:44:15 GMT</pubDate>
            <atom:updated>2026-08-17T15:44:15.313Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A learner’s take on one of the quietest, fastest-growing risks in software development</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*F2vgY3jmpYOHGqyE" /></figure><p>I’m still early in my journey into cloud and DevSecOps — no production pipeline of my own, no security team, just a lot of curiosity and a habit of going down rabbit holes. A few weeks ago, one of those rabbit holes led me to a term I hadn’t heard before: <strong>Shadow AI</strong>.</p><p>The more I read about it, the more it clicked. It’s not a futuristic threat. It’s already happening, quietly, inside almost every engineering org that uses AI tools — which, in 2026, is basically all of them.</p><p>Here’s what I learned, and why I think it deserves more attention than it’s getting.</p><h3>What Is Shadow AI, Exactly?</h3><p>“Shadow IT” is an old concept: employees using unapproved software or services without their company’s knowledge. Personal Dropbox accounts, unsanctioned Slack workspaces, that kind of thing.</p><p><strong>Shadow AI is the same idea, applied to AI tools</strong> — and it’s arguably more dangerous, because of what people paste into them.</p><p>Think about a normal day for a developer:</p><ul><li>Stuck on a bug at 11pm → pastes the function into a random AI chatbot to debug it</li><li>Needs test cases fast → pastes a whole module into an AI tool to generate them</li><li>Wants a code review before a PR → shares a chunk of the codebase with an AI assistant that isn’t sanctioned by the company</li></ul><p>None of this comes from bad intent. It comes from convenience. AI tools are genuinely useful, and reaching for the fastest one available is human nature.</p><p>The problem is what’s inside that code: <strong>API keys, database schemas, internal architecture details, customer data structures, or entire unreleased features.</strong> Once that’s pasted into an unapproved tool, it’s sitting on a server the company doesn’t control, governed by a terms-of-service nobody read, with no audit trail of what left the building.</p><h3>Why This Matters More Now</h3><p>A few things are converging to make this a bigger deal in 2026 than it was even a year ago:</p><p><strong>AI tools are everywhere.</strong> Nearly every developer now has access to multiple AI coding assistants, chatbots, and browser extensions — many free, many with no enterprise agreement behind them at all.</p><p><strong>Regulation has caught up.</strong> The EU AI Act’s high-risk provisions are now fully enforceable, which means how organizations handle AI-related data isn’t just a best-practice conversation anymore — it’s a compliance one, with real fines attached.</p><p><strong>Security teams can’t see what they can’t see.</strong> Traditional security tooling was built to monitor code repositories, endpoints, and network traffic — not what a developer pastes into a browser-based chat window. That gap is exactly where shadow AI lives.</p><p>Put together, you get a situation where the risk is growing, the stakes are rising, and most organizations still have no real visibility into it.</p><h3>Three Guardrails That Keep Coming Up</h3><p>As I read more, three practical mitigations kept showing up across different sources — not silver bullets, but a solid starting point.</p><h3>1. SBOM Checks</h3><p>A Software Bill of Materials (SBOM) is essentially an ingredients list for your software — every library, package, and dependency that went into it. The idea, applied to AI: <strong>treat AI-suggested code the same way you’d treat a third-party package.</strong> If an AI assistant suggests a dependency, it should show up on the radar and get vetted, not slip in silently because “the AI suggested it” felt like enough of a review.</p><h3>2. Secrets Scanning at the AI Boundary</h3><p>Most orgs already scan commits and repositories for secrets — API keys, credentials, tokens accidentally left in code. But by the time something reaches a commit, the leak may have already happened. The real point of exposure is often earlier: <strong>the moment someone pastes a </strong><strong>.env file or config block into a chat window.</strong> Catching it there, at the boundary between the developer and the AI tool, matters more than catching it three sprints later in a git-history audit.</p><h3>3. A Sanctioned-Tool List People Actually Want to Use</h3><p>This is the one that surprised me most. The instinct is to write a policy: “Don’t use unapproved AI tools.” But a policy nobody reads doesn’t stop shadow IT — it never has, for any category of shadow IT, ever. What actually works is giving developers a <strong>fast, approved alternative that’s genuinely good.</strong> People don’t route around good tools. They route around bad ones. If the sanctioned option is slower or worse than the free browser tool, developers will use the free browser tool — policy or not.</p><h3>What This Taught Me</h3><p>Coming from a learner’s perspective, this topic reframed how I think about security altogether. I used to picture cybersecurity as firewalls, network monitoring, and patching known vulnerabilities. Shadow AI made me realize security today is just as much about <strong>behavior and workflow</strong> as it is about infrastructure.</p><p>The tools people reach for out of convenience — not malice — are often the biggest blind spot. And that’s a much harder problem to solve with a firewall rule. It requires policy, tooling, and genuinely good alternatives working together.</p><h3>An Open Question</h3><p>I don’t have a production pipeline to test any of this on yet, so I’d genuinely like to hear from people who do: <strong>Is shadow AI something your team actively monitors? Or is it still flying under the radar, the way shadow IT did for years before anyone took it seriously?</strong></p><p>If you’re working in DevSecOps, cloud security, or platform engineering, I’d love to hear how your org is (or isn’t) handling this.</p><p><em>I’m currently learning DevSecOps and cloud security fundamentals and writing about what I find along the way. If you spot something I got wrong or oversimplified here, I’d genuinely appreciate the correction — that’s half the reason I’m writing this stuff down.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dd53226c7920" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[25 Years, 25 Lessons: Everything Life Handed Me Before I Was Ready]]></title>
            <link>https://medium.com/@poojaborade7022/25-years-25-lessons-everything-life-handed-me-before-i-was-ready-1ed5fe0216f5?source=rss-59a8239c7722------2</link>
            <guid isPermaLink="false">https://medium.com/p/1ed5fe0216f5</guid>
            <category><![CDATA[life-lessons]]></category>
            <category><![CDATA[personal-growth]]></category>
            <category><![CDATA[self-reflection]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[life]]></category>
            <dc:creator><![CDATA[Pooja Borade]]></dc:creator>
            <pubDate>Sat, 01 Aug 2026 20:56:09 GMT</pubDate>
            <atom:updated>2026-08-01T20:56:09.798Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>July 7, 2026. I turned 25.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9EuUU2gb2DpTdO9JWSxyjg.jpeg" /></figure><p>Twenty-five</p><p>It’s funny because when I was younger, 25 sounded so… <em>grown up</em>. I imagined I’d have every answer, a perfect career, a perfectly planned life, and somehow magically know what I was doing.</p><p>Spoiler alert: I don’t.</p><p>But I do have something much more valuable than all the answers — I have experiences.</p><p>Some made me laugh until my stomach hurt. Some made me cry when no one was watching. Some broke me, some built me, and some quietly changed the way I see the world forever.</p><p>I’m someone who laughs loudly, makes friends wherever I go, dreams a little too big, believes in impossible things, and loves my family more than anything. I grew up in a middle-class Indian family where values were never taught through lectures but through actions. Two years ago, I packed my bags, left home, and moved to the United States to chase a dream that once felt impossible.</p><p>These 25 lessons aren’t universal truths. They’re simply the things life whispered — and sometimes shouted — to me over the last quarter century.</p><p>If even one of them stays with you, this blog has done its job</p><p>Pour yourself a chai (or coffee, no judgment) — let’s get into it.</p><h3>1. Kindness is not optional</h3><p>My dad drilled this into me early: be kind, even to people who are rude or unkind to you. Kindness was never meant to be a transaction. It’s not about what someone deserves — it’s about who you choose to be.</p><h3>2. Hard work and faith are not competitors — they’re teammates</h3><p>Grind hard. Believe hard. Faith without effort is just wishing, and effort without faith is just exhausting. You need both hands on the wheel.</p><h3>3. Love hurts — and that’s exactly why it teaches you the most</h3><p>Loving someone deeply will crack you open in ways nothing else can. It’ll hurt in places you didn’t know existed. But it’ll also teach you things no book, no advice, no friend ever could.</p><h3>4. Your sibling is the one relationship that doesn’t need proving</h3><p>I have a sister, and I know — without a single doubt — that if the entire world turned its back on me tomorrow, she’d still be standing right there. That kind of certainty is rare. Protect it.</p><h3>5. Strength isn’t silence — I got that one wrong for years</h3><p>Since childhood, I buried my emotions instead of showing them. I thought that made me strong. It didn’t. It just built walls so high that now, even the people who love me can’t get past them. If I could rewrite one habit, it’d be this: say the feeling out loud. Don’t let silence pass as strength.</p><h3>6. Being a good listener doesn’t mean someone is listening back</h3><p>I’ve always been the friend, the daughter, even the stranger who listens. But outside of my sister, I’m still looking for someone who listens to <em>me</em> the way I listen to everyone else. And that’s okay to admit.</p><h3>7. Family is life’s greatest investment.</h3><p>Jobs come and go.<br>Money comes and goes.<br>Success comes and goes.<br>Family stays.<br>Celebrate them while they’re here.<br>Call them.<br>Visit them.<br>Love them loudly.</p><h3>8. Cherish every single moment like there’s no tomorrow</h3><p>Because honestly? There might not be. Stop waiting for birthdays. Stop waiting for promotions. Stop waiting for “someday.” Life is happening today. Celebrate random Tuesdays. Dance in your kitchen. Laugh too much. Take pictures. Stop saving the good moments for “later.”</p><h3>9. Live a little before “adulting” takes the wheel</h3><p>Travel. Eat the junk food. Do the reckless, nonsensical, twenty-something things — before the job comes and forces you to choose between your time and your paycheck. Trust me, that trade gets harder, not easier.</p><h3>10. “I don’t have time” is almost always a choice, not a fact</h3><p>I’ve watched friends say they’re too busy to call their own parents. Meanwhile, I’ve lived in the US for two years and haven’t missed a single day calling mine. You always find time for what actually matters to you.</p><h3>11. Life laughs at your five-year plan — let it</h3><p>I used to imagine every detail of my sister’s wedding: the outfit, the dance, the moment. Then life happened. I was living abroad when she got married, and I sat in the crowd as a guest instead of standing beside her the way I’d pictured. Lesson learned — stop scripting every moment. Let life surprise you instead of punishing you for not following the plan.</p><h3>12. Bold moves scare you right before they change your life</h3><p>I come from a middle-class family with no roadmap for what I wanted. Today, I hold a master’s degree from the US because I stopped waiting for the fear to leave and moved anyway. Don’t let where you started convince you of where you can’t go.</p><h3>13. Outgrowing people isn’t betrayal — it’s just growth</h3><p>You won’t feel the same about everyone you’ve known for years, and that’s not a character flaw. People evolve. So do you. Let that be okay.</p><h3>14. Strangers sometimes see you clearer than the people who’ve known you for years</h3><p>Some of the most meaningful conversations of my life happened with people I never expected to matter — and never saw coming. Don’t close that door.</p><h3>15. Never forget your roots</h3><p>When I left India, my dad handed me a few Indian currency notes.He smiled and said,</p><p><em>“Keep this in your wallet. Never forget where you came from.”</em></p><p>I still carry them. I always will.</p><h3>16. People change even if your family doesn’t</h3><p>When I went back home for vacation, my family and my sister felt exactly the same. But relatives, friends — some of them had shifted in how they treated me. It stung for a second. Then I learned: don’t let that change you.</p><h3>17. Chase your dreams like there’s no other life waiting for you</h3><p>Struggle for it. Dreams really do come true — but usually only for the people who refuse to sit still.</p><h3>18. Wear the pretty dress today</h3><p>Even if you’re just going out to buy a bread packet — get dressed up if you want to. Don’t save that outfit hanging in your cupboard for “the right time.” You’re not promised one.</p><h3>19. Friendships fade, and that’s part of growing up</h3><p>As a child, I thought friends stayed forever.</p><p>At 25, I can count my closest friends on one hand.</p><p>And that’s okay.</p><p>Some people stay for chapters. Some stay for the whole book.</p><p>Both matter.</p><h3>20. Be the first one to reach out</h3><p>Don’t be afraid to text first, call first, show up first. Someone has to break the silence — let it be you.</p><h3>21. Ego kills relationships faster than distance ever could</h3><p>Communication is everything. I’ve watched my own relationships get damaged not because love ran out, but because ego walked in. Talk it out before you let silence do the damage.</p><h3>22. Be real. Social media is a highlight reel, not a life</h3><p>Don’t perform a version of yourself you can’t actually afford. Authenticity will always age better than a curated feed.</p><h3>23. Loyalty is a flex, even when the other person doesn’t deserve it</h3><p>I’ve stayed loyal in every friendship and relationship I’ve ever been in. I got cheated on once. But looking back now, I’m proud — because staying true was never about them. It was about who I decided to be.</p><h3>24. A little selfishness isn’t a crime, it’s self-preservation</h3><p>Put yourself first sometimes. That doesn’t mean abandoning the people you love — it means remembering you’re on that list too.</p><h3>25. Learn the art of ignoring.</h3><p>This might be my favorite lesson. Ignore unnecessary opinions. Ignore gossip. Ignore negativity. Ignore people determined to misunderstand you.</p><p>Whenever someone speaks badly about me, I remind myself of one thing:</p><p><strong>God is watching everything.</strong></p><p>I don’t need to prove myself to everyone. Life has its own way of balancing things. And that’s enough peace for me.</p><h3>Here’s to the Next 25, May they be even louder.</h3><p>I’m still figuring life out.</p><p>I still make mistakes.</p><p>I still overthink.</p><p>I still cry.</p><p>I still dream too much.</p><p>But if these first 25 years have taught me anything, it’s this:</p><p>Life isn’t about having everything figured out.</p><p>It’s about becoming someone you’re proud of while figuring it out.</p><p>Here’s to more dreams.</p><p>More countries.</p><p>More friendships.</p><p>More family dinners.</p><p>More courage.</p><p>More kindness.</p><p>And hopefully…</p><p>Many more lessons worth writing about.</p><p><strong>Happy 25th to me. ❤️</strong></p><p><em>If you’ve read this far, thank you for being a small part of my journey. I’d love to know — what’s one life lesson that changed you forever? Let’s talk in the comments.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1ed5fe0216f5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Perfectionism: A One-Woman Show That Never Opens]]></title>
            <link>https://medium.com/@poojaborade7022/perfectionism-a-one-woman-show-that-never-opens-8a56be9d522e?source=rss-59a8239c7722------2</link>
            <guid isPermaLink="false">https://medium.com/p/8a56be9d522e</guid>
            <category><![CDATA[self-improvement]]></category>
            <category><![CDATA[perfectionism]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[motivation]]></category>
            <category><![CDATA[personal-development]]></category>
            <dc:creator><![CDATA[Pooja Borade]]></dc:creator>
            <pubDate>Fri, 24 Jul 2026 01:21:21 GMT</pubDate>
            <atom:updated>2026-07-24T01:21:21.736Z</atom:updated>
            <content:encoded><![CDATA[<p>Confessions of a recovering perfectionist, and the three things that finally got me to stop.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*mPAYw_jFHUYx35FF.jpg" /></figure><p>Let me tell you about a show I’ve been producing, directing, and starring in for years. It’s called <em>The Perfect Me</em>.</p><p>She wakes up at 5 a.m. Her inbox is always at zero. Her reports never have a typo. She’s never said “um” mid-sentence, never sent an email and spotted the mistake three seconds too late, never once shown up to something feeling underprepared.</p><p>I have spent years rehearsing for her opening night.</p><p>It has never once opened.</p><p>Every time the curtain got close to rising, I found one more thing to fix. One more line to polish. One more pass “just to be safe.” Meanwhile, the audience — my career, my confidence, my actual life — sat in the dark, waiting for a show that was always “almost ready.”</p><p>Here’s the thing I eventually learned, the hard and slightly embarrassing way: perfection isn’t a finish line. It’s a horizon. It moves every time you get close to it. You can spend your whole life walking toward it and never arrive, because it was never designed to be arrived at.</p><p>So this isn’t a “just believe in yourself” post. It’s the three specific things that actually got my curtain up.</p><h3>Stop looking inward. Start looking sideways.</h3><p>Perfectionism has a nasty trick: it convinces you that you’re the only one who isn’t ready yet. Everyone else, it whispers, has it figured out.</p><p>The fix wasn’t more self-belief. It was better evidence. I started paying closer attention to the people around me — the colleague who sent the “good enough” version of a report and got promoted anyway, the speaker who forgot a line on stage and got a laugh and applause instead of pity. Every one of them was proof of the same thing: nobody was actually waiting for flawless. They were waiting for <em>real</em>, and real was already working for everyone but me.</p><p>This is sometimes called social proof, and it’s oddly one of the fastest cures for perfectionism I’ve found. You don’t need to trust your own judgment about what’s “ready.” You just need to borrow the evidence sitting right next to you.</p><h3>Shrink the finish line until it’s laughably small.</h3><p>“Write the perfect report” is not a task. It’s a life sentence. Vague, enormous goals like that don’t motivate — they paralyze, because your brain can’t find a door into them.</p><p>So I stopped setting goals like that entirely. Instead: “Write one honest paragraph before lunch.” “Get through the first three sentences without stopping to fix them.” Small, specific, almost insultingly achievable.</p><p>It felt silly at first — like setting the bar so low I was basically stepping over it lying down. But momentum doesn’t care about your dignity. It just needs a door small enough to walk through. Once I was moving, the rest of the task usually finished itself.</p><h3>Make the cost of waiting louder than the fear of failing.</h3><p>This is the one that actually broke the spell.</p><p>I stopped asking, “What if I fail?” and started asking, “What am I <em>already</em> losing by not trying?”</p><p>The report I sat on for one more day of polishing was a day someone else used to ship theirs first. The idea I didn’t say out loud in a meeting because it wasn’t phrased perfectly yet was an idea someone else said — clumsily — and got the credit for anyway. Perfectionism had convinced me the only risk on the table was embarrassment. It never once mentioned the much bigger one: irrelevance.</p><p>Once I started counting <em>that</em> cost instead, “not ready yet” stopped sounding responsible and started sounding expensive.</p><h3>The show that finally opened</h3><p><em>The Perfect Me</em> is never going to open. Turns out that’s fine — her reviews would’ve been terrible anyway. Nobody actually wants to watch someone who never takes a risk; there’s no story in flawless.</p><p>The show that did open was smaller, messier, and infinitely more interesting: <em>The Good Enough Me, Doing It Anyway.</em> She sends the email before rereading it a fifth time. She raises her hand with a half-formed idea. She gives the talk with the shaky hands.</p><p>If you’re backstage right now, in your own version of a show that never opens — this is your nudge. Borrow someone else’s courage. Shrink the finish line. Count what waiting is actually costing you.</p><p>Then, however small the step — open it.</p><p><em>If this resonated, I’d love to hear the show you’ve been keeping in rehearsal. Drop it in the comments — sometimes just saying it out loud is the first step toward opening night.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8a56be9d522e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Meet Sol, Terra, and Luna: OpenAI Just Renamed the Future, and Almost Nobody Can Use It Yet]]></title>
            <link>https://medium.com/@poojaborade7022/meet-sol-terra-and-luna-openai-just-renamed-the-future-and-almost-nobody-can-use-it-yet-bad1ab23f9f6?source=rss-59a8239c7722------2</link>
            <guid isPermaLink="false">https://medium.com/p/bad1ab23f9f6</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[future-of-ai]]></category>
            <category><![CDATA[ai-tools]]></category>
            <category><![CDATA[technews]]></category>
            <dc:creator><![CDATA[Pooja Borade]]></dc:creator>
            <pubDate>Mon, 06 Jul 2026 01:53:07 GMT</pubDate>
            <atom:updated>2026-07-06T01:53:07.128Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*L0Dz2hqDzSciO4vO.png" /></figure><p>OpenAI just did something it hasn’t done in a while: it launched a model family that most of us can’t actually touch. GPT-5.6 arrived this week as three distinct models — Sol, Terra, and Luna and while the benchmarks are genuinely impressive, the launch itself tells a bigger story than the numbers do. This is what’s coming. Here’s what you need to know before it lands in your hands.</p><h3>Goodbye “Mini,” Hello Cosmos</h3><p>First, the naming. OpenAI is retiring the “nano” and “mini” suffixes that have defined its model lineup for years. In their place: Sol, Terra, and Luna names meant to represent durable capability tiers rather than shrunken-down versions of a flagship. As OpenAI put it, the generation number tells you when a model was made, while the name tells you what it’s built for. Sol, Terra, and Luna can each advance on their own schedule going forward, instead of being locked to a single release cycle.</p><p>It’s a small branding shift, but it signals something real: OpenAI no longer thinks of “smaller” models as compromises. They’re distinct products with their own reasons to exist.</p><h3>The Three Models, Broken Down</h3><p><strong>Sol</strong> is the flagship — the model built for the hardest problems: long coding sessions, complex multi-step reasoning, and security-focused work. It’s also the only one of the three that unlocks two new modes: a “max” reasoning setting for deeper thinking, and an “ultra” mode that splits a task across multiple sub-agents working in parallel. Sol is where OpenAI is putting its biggest capability claims, including a new state-of-the-art result on Terminal-Bench, a benchmark that tests how well a model can operate a command line to complete real, multi-step engineering tasks. Priced at $5 per million input tokens and $30 per million output tokens, it costs the same as the previous flagship — but OpenAI says it delivers meaningfully more for that price, especially on long-running coding and agentic work.</p><p><strong>Terra</strong> is the one worth paying closest attention to. It’s positioned as the new “default” — the everyday-work model — and it reportedly matches the performance of the previous flagship model while costing about half as much. That’s the pattern to watch across the AI industry right now: last generation’s top-tier performance keeps sliding down into the mid-tier price bracket, generation after generation. If that holds, it means the “good enough for everything” model is getting dramatically cheaper roughly every few months.</p><p><strong>Luna</strong> is the speed-and-budget option, aimed at high-volume tasks like classification, autocomplete, or routing — the unglamorous, high-frequency work that a business might run millions of times a day. It’s the cheapest of the three, and OpenAI is careful to note that cheapest doesn’t mean weakest across the board; on several tasks, Luna reportedly holds its own.</p><p>OpenAI is also rolling out a Cerebras-hosted version of Sol this month, targeting inference speeds of up to 750 tokens per second — aimed squarely at use cases where waiting for a response is the actual bottleneck, not model quality.</p><h3>Here’s the Catch: You Probably Can’t Use It Yet</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JW-H7GxtsXDwfzP5" /></figure><p>This is the part that makes GPT-5.6 different from a typical OpenAI launch. It is not available in ChatGPT. It’s not available to the general API either. Access, for now, is limited to roughly twenty pre-approved organizations, reachable only through the API or OpenAI’s coding tool, Codex.</p><p>Why the narrow rollout? OpenAI says it briefed the U.S. government on the model’s capabilities ahead of launch and, at the government’s request, is starting with a small, vetted group before expanding. That’s tied to a recent executive order directing federal agencies to review new frontier models for safety and national-security implications before they reach the public. OpenAI has been explicit that it doesn’t view this kind of staged, government-coordinated release as an ideal long-term setup — but for now, it’s the price of shipping a model with a meaningful jump in cybersecurity capability.</p><p>That capability jump is exactly why the caution exists. OpenAI’s own safety testing found Sol and Terra can identify vulnerabilities and even pull together pieces of an exploit, but couldn’t independently execute a complete attack against hardened systems in testing. The company frames this as a net positive for defenders — security teams doing code review, patching, and vulnerability research get a serious upgrade — while the safeguards are designed to make misuse harder and more detectable.</p><h3>What “Coming Soon” Actually Means Here</h3><p>OpenAI says general availability is planned “in the coming weeks,” with no firm date attached. There are also early hints — spotted in a research paper about genomics benchmarks — that ChatGPT Pro subscribers might eventually get access to extended “Pro” versions of all three models, not just the single best model as has been the case historically. Nothing here is confirmed as a product yet, but it suggests OpenAI is thinking about giving Pro users a choice of how a model thinks, not just access to the smartest one.</p><h3>Why This Matters Even If You Can’t Use It Today</h3><p>It’s tempting to shrug off a model you can’t actually try. Don’t. Three things about this launch are worth internalizing right now, regardless of when Sol, Terra, and Luna show up in your ChatGPT dropdown:</p><p><strong>The price-to-performance curve keeps bending down.</strong> Terra matching a previous flagship at half the price is the headline that actually affects your wallet, whether you’re an individual builder or running a team on API costs.</p><p><strong>Frontier AI releases are becoming a government-coordinated process, not just a product launch.</strong> This is likely not a one-off. As models get better at cybersecurity and other high-stakes domains, expect more releases to come with a “preview to trusted partners first” phase before the public ever sees them.</p><p><strong>“Reasoning” and “agentic” are no longer premium add-ons — they’re standard.</strong> The old idea of a separate “thinking” model bolted onto a base model is gone. Reasoning is just baked into the main model now, and sub-agent collaboration (OpenAI’s “ultra” mode) is becoming a built-in feature rather than something you’d have to engineer yourself.</p><h3>The Bottom Line</h3><p>GPT-5.6 is a preview of where the ground is shifting, even if most of us are stuck watching from the sidelines for a few more weeks. The intelligence is getting cheaper, the capabilities are getting sharper, and the rollout process is getting more cautious, all at the same time. That combination — faster progress, tighter gatekeeping — might end up being the real story of AI in the second half of 2026, not any single model’s benchmark score.</p><p>Keep an eye on your model picker. Sol, Terra, and Luna are coming — the only real question left is when.</p><p><em>Would you want your AI tools to go through a government safety check before you get access, or does that feel like it’s slowing down innovation? Let me know what you think in the comments.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bad1ab23f9f6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI So Powerful, Its Own Creators Won’t Let You Use It]]></title>
            <link>https://medium.com/@poojaborade7022/the-ai-so-powerful-its-own-creators-wont-let-you-use-it-8c909175c243?source=rss-59a8239c7722------2</link>
            <guid isPermaLink="false">https://medium.com/p/8c909175c243</guid>
            <category><![CDATA[future-of-tech]]></category>
            <category><![CDATA[mytho]]></category>
            <category><![CDATA[project-glasswing]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[claude]]></category>
            <dc:creator><![CDATA[Pooja Borade]]></dc:creator>
            <pubDate>Mon, 15 Jun 2026 18:38:02 GMT</pubDate>
            <atom:updated>2026-06-15T18:38:02.224Z</atom:updated>
            <content:encoded><![CDATA[<h3>Inside Claude Mythos, Project Glasswing, and the strange new calculus of “too capable to release”</h3><p>Somewhere in a server room, an AI model is currently scanning through the source code that runs your bank, your hospital, and the power grid that keeps your lights on. It’s hunting for flaws — the kind that hackers spend months searching for. And it’s finding thousands of them.</p><p>You can’t use this AI. You can’t sign up for it. You can’t even put your name on a waitlist for it.</p><p>Not because it’s broken. Because it works <em>too well</em>.</p><p>This is the story of Claude Mythos, the most capable AI model Anthropic has ever built — and the first one they deliberately chose to keep locked away.</p><h3>The model that broke the playbook</h3><p>For years, the AI release playbook has been remarkably consistent: a lab builds a more capable model, runs some safety evaluations, and then ships it. Maybe there’s a staged rollout, maybe some red-teaming, but the destination is always the same — broad availability. Developers get API access. Consumers get an app. The model becomes a product.</p><p>In April 2026, Anthropic broke that pattern. They announced Mythos Preview, described as their most capable model yet, and then said something no major AI lab had said quite so plainly before: they were not making it generally available.</p><p>The reason wasn’t underperformance. It was the opposite. Mythos had become extraordinarily good at one particular skill — finding security vulnerabilities in software, quickly, and at a scale that outpaced what teams of human researchers could do.</p><p>That single capability created a problem nobody had fully grappled with before.</p><h3>The same flashlight, pointed two ways</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*99AF2vu652013-kMVKUshQ.jpeg" /></figure><p>Here’s the uncomfortable truth at the center of this story: the skill that makes an AI valuable for <em>defense</em> is identical to the skill that makes it dangerous as a <em>weapon</em>.</p><p>A model that can scan a hospital’s codebase and say “here are the twelve places an attacker could get in” is performing exactly the same task as a model that could scan that same codebase and hand those twelve entry points to someone with bad intentions. There’s no technical difference between finding a hole to patch it and finding a hole to walk through it.</p><p>This is the dual-use dilemma, and it’s not new to AI — it’s the same logic that governs research into pathogens, or encryption, or any technology that can protect or harm depending on who’s holding it. What’s new is the <em>speed and scale</em>. A human security researcher might find a handful of critical vulnerabilities in a major piece of software over months of work. Mythos-class models can plausibly do this continuously, across thousands of codebases, simultaneously.</p><p>So Anthropic faced a choice: release it broadly and hope the defenders move faster than the attackers, or hold it back and try to give defenders a head start before this capability becomes common.</p><p>They chose the second option. The result is called Project Glasswing.</p><h3>What Project Glasswing actually is</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Sqv94uaJqCq8_u1y8mNHKQ.jpeg" /></figure><p>Project Glasswing is, in essence, a controlled trial run for living in a world where AI can find software vulnerabilities faster than humans can patch them.</p><p>In April 2026, Anthropic gave roughly 50 organizations early access to Mythos — not to build products with it, but to point it at the world’s most critical software and start fixing what it found. The list of participants reads like a who’s-who of the institutions modern life depends on: major cloud providers, hardware manufacturers, financial institutions, open-source foundations, and government agencies.</p><p>The results, even in the first few weeks, were striking. Partners reportedly uncovered more than ten thousand high- or critical-severity vulnerabilities across systemically important software — flaws that, left unpatched, could have been discovered and exploited by someone with very different intentions.</p><p>By June 2026, just two months later, the program had nearly tripled — expanding to around 150 organizations spanning more than 15 countries. Crucially, the expansion specifically targeted sectors that hadn’t been well represented in the first round: power utilities, water systems, healthcare providers, and telecommunications companies. The logic was explicit — these are the sectors where a successful attack doesn’t just cost money, it can affect the daily life of a hundred million people or more.</p><h3>The new bottleneck nobody saw coming</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aP4QwLh1Rh1K2fOtedepBw.jpeg" /></figure><p>For decades, the limiting factor in cybersecurity has been <em>discovery</em>. Finding vulnerabilities was slow, expensive, and required scarce, highly skilled human talent. Patching was the easy part, comparatively — once you knew where the hole was, fixing it was often straightforward.</p><p>Glasswing flips that equation entirely.</p><p>With a model like Mythos doing the discovery work, the volume of vulnerabilities found can outpace the human capacity to verify, prioritize, disclose, and patch them. The bottleneck isn’t finding the cracks anymore. It’s everything that has to happen <em>after</em> you find them.</p><p>This is, in a way, a good problem to have — it means defenders are getting an unprecedented advantage. But it also means the entire cybersecurity industry has to restructure around a new reality: AI can generate more actionable security intelligence in a week than most organizations can act on in a year.</p><h3>The warning embedded in the announcement</h3><p>Here’s the part of this story that deserves the most attention, and it’s easy to miss because it’s stated so calmly.</p><p>Anthropic has openly said that within six to twelve months, other AI companies will likely have models with similar capabilities to Mythos — and that those companies might not build in the same safeguards, or choose the same cautious release strategy.</p><p>Read that again. The company that built this model isn’t claiming they’ve solved the problem. They’re saying they’ve bought <em>time</em> — a window in which defenders can get ahead, before this kind of capability becomes commoditized and potentially available without restriction.</p><p>Project Glasswing, in other words, isn’t really the destination. It’s a rehearsal. It’s the AI industry’s attempt to answer a question that’s about to become unavoidable: what happens when the ability to find — and exploit — software vulnerabilities at scale becomes cheap, fast, and widely accessible?</p><h3>The questions we can’t avoid</h3><p>This story raises questions that don’t have comfortable answers, and they’re worth sitting with.</p><p><strong>Should the most capable AI models ever be released to the public at all?</strong> For years, the assumption in AI has been that broader access is generally good — it democratizes capability, accelerates innovation, and avoids concentrating power. Mythos is a case where that assumption visibly breaks down. But if the most powerful models are withheld, who decides which models cross that line, and based on what evidence?</p><p><strong>Who gets early access when a model is “too powerful” for the public?</strong> The Glasswing partner list is dominated by large corporations and governments — the institutions that already have significant resources and influence. That may be the correct call given what’s at stake. But it’s worth noticing that “too dangerous for the public” and “available to a select group of powerful organizations” are, in practice, the same sentence.</p><p><strong>Is this responsible caution, or an early sketch of how AI power gets distributed going forward?</strong> Maybe both. Maybe neither framing is quite right. But the pattern — a capability emerges, it’s deemed too significant for open release, and access becomes the province of a privileged few — is one we should expect to see again, and probably more often, as models keep getting more capable.</p><h3>Why this isn’t a “someday” story</h3><p>It’s tempting to file stories like this under speculative future-of-AI content — interesting, but not urgent. That would be a mistake.</p><p>Project Glasswing isn’t a thought experiment or a policy paper. It’s an active program, running right now, with real organizations using a real model to find real vulnerabilities in the software that the modern world runs on. The model finding those vulnerabilities exists today. The 10,000-plus flaws it found are real flaws that existed in real systems.</p><p>And the warning that other companies will have similar models within months isn’t a hedge — it’s a prediction made by the people who would know best, because they’re the ones who just crossed this threshold first.</p><p>We are, right now, in the early days of a world where AI doesn’t just generate text, images, or code — it <em>finds things</em>. It finds the weak points in the systems we depend on. And the question of who controls that capability, and how it gets used, is no longer hypothetical.</p><p>The age of AI as a creative tool is well underway. The age of AI as a tool for finding — and exploiting, or defending — the hidden seams in our digital infrastructure has just begun. Project Glasswing is our first real look at what that age might look like.</p><p>It’s worth paying attention.</p><p><em>Sources: Anthropic’s public announcements on Claude Mythos Preview and Project Glasswing (April–June 2026), and reporting from TechCrunch, Cybersecurity Dive, and Schneier on Security covering the program’s expansion.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8c909175c243" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>