<?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 Ege Kaan Işık on Medium]]></title>
        <description><![CDATA[Stories by Ege Kaan Işık on Medium]]></description>
        <link>https://medium.com/@egekaanisik?source=rss-ca2093f71a1d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*vh-ke_QVpqFP9yYbzYvKuA.jpeg</url>
            <title>Stories by Ege Kaan Işık on Medium</title>
            <link>https://medium.com/@egekaanisik?source=rss-ca2093f71a1d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 00:52:13 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@egekaanisik/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[⚡ LiteRT-LM: How Google’s New Edge Engine Changes In-Browser LLM Inference]]></title>
            <link>https://medium.com/dogus-teknoloji/litert-lm-how-googles-new-edge-engine-changes-in-browser-llm-inference-ebce8e56324f?source=rss-ca2093f71a1d------2</link>
            <guid isPermaLink="false">https://medium.com/p/ebce8e56324f</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai-web]]></category>
            <category><![CDATA[javascript]]></category>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[litert-lm]]></category>
            <dc:creator><![CDATA[Ege Kaan Işık]]></dc:creator>
            <pubDate>Tue, 29 Sep 2026 05:04:20 GMT</pubDate>
            <atom:updated>2026-09-30T11:13:08.213Z</atom:updated>
            <content:encoded><![CDATA[<h4>For years, delivering generative AI in web apps followed a rigid playbook: spin up expensive GPU clusters in the cloud, gate them behind authentication, and stream responses back to client browsers token by token.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SpbZnoZ3zopeO6QHNWHzjA.png" /></figure><p>While this cloud-centric model launched the generative AI era, it introduced distinct trade-offs: continuous server expenses, network latency, and privacy hurdles when handling sensitive user documents.</p><p>In-browser AI execution emerged to challenge that model. Early prototypes using WebGPU and WebAssembly proved local browser inference was possible, but integrating these workflows often felt brittle. Now, Google has introduced <strong>LiteRT-LM</strong>.</p><p>As Google evolved TensorFlow Lite into <strong>LiteRT</strong>, LiteRT-LM was released as a specialized inference framework built specifically for language models across edge platforms, including the browser.</p><p>What makes LiteRT-LM distinct from predecessors like MediaPipe’s LLM Inference API, and how does it reshape the developer experience? Let’s take a closer look.</p><h3>The Predecessor: MediaPipe’s LLM Inference API</h3><p>To understand where LiteRT-LM fits, it helps to examine its direct predecessor in the Google ecosystem: <strong>MediaPipe Tasks GenAI</strong> (and its LlmInference API).</p><p>MediaPipe was originally architected for classic computer vision and audio pipelines, such as background blur, hand tracking, and face landmark detection. When Google brought language models like Gemma to the browser, they packaged it using that existing design pattern: as a self-contained <strong>Task</strong>.</p><p>The developer pattern looked like this:</p><pre>import { FilesetResolver, LlmInference } from &#39;@mediapipe/tasks-genai&#39;;<br><br>const genai = await FilesetResolver.forGenAiTasks(&#39;wasm-url&#39;);<br>const llm = await LlmInference.createFromModelPath(<br>  genai,<br>  &#39;/models/gemma-4-E2B-it-web.litertlm&#39;<br>);<br><br>const result = await llm.generateResponse(&quot;Summarize this document...&quot;);</pre><p>While this was a milestone for web-based machine learning, developers building production chat interfaces or agents quickly encountered structural hurdles:</p><ul><li><strong>Task-Oriented vs. Stateful Conversational:</strong> MediaPipe primarily treated the model as a stateless function: you send input text, and it returns a string. Building an interactive multi-turn conversation meant manually formatting system prompts, re-passing the historical dialogue, and managing prompt formatting tokens (like &lt;start_of_turn&gt;user...) yourself.</li><li><strong>Rigid Memory Management:</strong> A language model needs to remember earlier turns in a conversation using something called a <strong>KV-cache</strong> (Key-Value cache), which stores contextual memory. Under early task abstractions, re-evaluating conversations or keeping multiple chat tabs open could bloat browser tab memory or require recalculating long prompts from scratch.</li><li><strong>Brittle Lifecycle Control:</strong> Interrupting an ongoing response, pruning conversational tokens, or cloning a session for multiple parallel workflows required substantial custom wrapper code.</li></ul><h3>The Shift: LiteRT-LM’s Runtime Architecture</h3><p>LiteRT-LM was developed to tackle these constraints directly. Rather than treating an LLM like an image filter, it treats the browser as a true <strong>conversational runtime</strong>.</p><p>LiteRT-LM decouples the heavy model from individual interactions:</p><ol><li><strong>The Engine:</strong> Loads the heavy, immutable model weights into WebGPU buffers once.</li><li><strong>The Conversation Session:</strong> A lightweight state object that manages the KV-cache, chat roles, and message history.</li></ol><pre>import { Engine } from &#39;@google-ai-edge/litert-lm&#39;;<br><br>// 1. Initialize the shared engine (weights stay in GPU memory once)<br>const engine = await Engine.create({<br>  model: &#39;/models/gemma-4-E2B-it-web.litertlm&#39;<br>});<br><br>// 2. Spawn an isolated, stateful conversation<br>const chat = await engine.createConversation({<br>  preface: {<br>    messages: [{ role: &#39;system&#39;, content: &#39;You are a private offline writing assistant.&#39; }]<br>  }<br>});<br><br>// 3. Native streaming via standard JavaScript Async Iterables<br>for await (const chunk of chat.sendMessageStreaming(&quot;Review this paragraph...&quot;)) {<br>  updateUI(chunk);<br>}</pre><h3>Core Differences: LiteRT-LM vs. MediaPipe LLM Inference</h3><p>The architectural divide between the two approaches boils down to a few critical design choices:</p><h4>1. Underlying Philosophy</h4><ul><li><strong>MediaPipe:</strong> Task-centric. It approaches generative text as an isolated ML task to run once and finish.</li><li><strong>LiteRT-LM:</strong> Runtime-centric. It operates as an ongoing host environment that maintains conversational state inside your web app.</li></ul><h4>2. Multi-Turn Conversation State</h4><ul><li><strong>MediaPipe:</strong> Manual and stateless. Developers must manually stitch prompt templates, track chat history, and append special control tokens for each conversational turn.</li><li><strong>LiteRT-LM:</strong> Native and stateful. The createConversation API handles message histories, system role prefaces, and prompt formatting automatically.</li></ul><h4>3. Memory &amp; KV-Cache Management</h4><ul><li><strong>MediaPipe:</strong> Context must often be re-processed across long interactions, which increases processing time and memory pressure.</li><li><strong>LiteRT-LM:</strong> Explicit KV-cache lifecycle management. It retains prior conversation state in GPU memory and provides fine-grained controls over context budgets.</li></ul><h4>4. Session Branching</h4><ul><li><strong>MediaPipe:</strong> Complex. Creating an alternate branch of an ongoing chat requires feeding the full history into a fresh task instance.</li><li><strong>LiteRT-LM:</strong> Supported out of the box. Sessions can be cloned or branched without recomputing the common context prefix from scratch.</li></ul><h4>5. Streaming and Lifecycle Hooks</h4><ul><li><strong>MediaPipe:</strong> Relies primarily on callback patterns with limited abort and cleanup mechanisms.</li><li><strong>LiteRT-LM:</strong> Embraces modern JavaScript async iterables (for await...of) alongside clean abort handlers like cancel() and memory disposal via delete().</li></ul><p>In short: <strong>MediaPipe asks, <em>“Do you want to run this generative task?”</em> while LiteRT-LM asks, <em>“Do you want to host an interactive model inside your web app?”</em></strong></p><h3>Why In-Browser Inference Matters for Real-World Apps</h3><p>Running language models inside browser memory alters web architecture across three practical areas:</p><h4>Zero Cloud Inference Bills</h4><p>Cloud-hosted LLM APIs bill for every token generated. For free tiers or applications with thousands of active users, these compute expenses scale rapidly. In an in-browser model, your static hosting serves the model asset once. After caching, user interactions consume zero server compute.</p><h4>Built-in Privacy and Compliance</h4><p>When analyzing user data (such as medical notes, legal documents, or proprietary source code), routing text through third-party servers often requires strict legal and data-handling review. With in-browser execution, the user’s data remains isolated within their local browser sandbox.</p><h4>True Offline Capability (Zero Network Latency)</h4><p>Once weights are stored locally, apps built with service workers or Progressive Web App (PWA) architectures remain fully functional without internet access. Because inference runs over high-speed internal GPU buses rather than public HTTP roundtrips, response times bypass queueing and network latency.</p><h3>Engineering Realities: The Elephant in the Browser</h3><p>Building with local browser inference requires navigating real-world constraints:</p><h4>The Initial Download</h4><p>A quantized, web-optimized model like gemma-4-E2B-it-web still ranges around 1 to 2 gigabytes. Loading that asset via standard fetch() into a single JavaScript memory buffer can trigger tab memory crashes.</p><p>Production implementations typically stream the model download directly into the <strong>Origin Private File System (OPFS)</strong>, a specialized, high-performance browser storage API. This approach writes raw chunks to disk during download, allowing the application to initialize the model without exhausting main-thread memory.</p><h4>Heterogeneous Client Hardware</h4><p>On the server, developers can count on high-end hardware like an NVIDIA H100 or A100. On the web, users run diverse environments: some have dedicated desktop GPUs, while others rely on shared system memory on low-power laptops. As a result, token speeds vary based on the host device’s available memory bandwidth.</p><h3>Looking Forward</h3><p>In-browser language modeling has shifted past early proof-of-concept experiments.</p><p>By replacing low-level pipelines with a dedicated, stateful conversational runtime, <strong>LiteRT-LM</strong> provides the stability needed to build reliable local AI experiences. Instead of treating local AI as a novelty, web developers can treat the client browser as a first-class compute engine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ebce8e56324f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/dogus-teknoloji/litert-lm-how-googles-new-edge-engine-changes-in-browser-llm-inference-ebce8e56324f">⚡ LiteRT-LM: How Google’s New Edge Engine Changes In-Browser LLM Inference</a> was originally published in <a href="https://medium.com/dogus-teknoloji">Doğuş Teknoloji</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ Fixing AI-Generated Legacy Code: Inside the “Modern Web Guidance” Initiative]]></title>
            <link>https://medium.com/dogus-teknoloji/fixing-ai-generated-legacy-code-inside-the-modern-web-guidance-initiative-a3e173407d6e?source=rss-ca2093f71a1d------2</link>
            <guid isPermaLink="false">https://medium.com/p/a3e173407d6e</guid>
            <category><![CDATA[ai-web]]></category>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[modern-web-guidance]]></category>
            <category><![CDATA[ai-coding]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Ege Kaan Işık]]></dc:creator>
            <pubDate>Thu, 16 Jul 2026 22:27:03 GMT</pubDate>
            <atom:updated>2026-09-30T11:09:51.827Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*k5I4OyyqL1Sj_hK_QhlHww.jpeg" /></figure><p><em>Why AI-assisted development defaults to outdated patterns, and how a collaborative cross-browser standard is bringing the modern web platform to your IDE.</em></p><p>If you have used AI coding assistants like Claude, Copilot, or Gemini recently, you have likely felt the thrill of generating an entire user interface in seconds. But if you analyze the code your AI assistant produces, you will notice a frustrating trend.</p><p>Despite all their computational power, <strong>AI coding agents consistently default to older, outdated solutions.</strong></p><p>They will readily write fifty lines of complex, custom JavaScript to build a scroll animation, handle an accordion, or establish positioning for a tooltip. Meanwhile, modern CSS and HTML already have native, high-performance APIs designed to handle those exact problems out of the box.</p><p>Why does this happen, and how can we fix it? The answer lies in a major cross-browser collaboration supported by the Google Chrome team, the Microsoft Edge team, and the web development community: <strong>Modern Web Guidance</strong>.</p><h3>The AI Training Gap: Why Agents Write Bloated Code</h3><p>AI models do not write legacy code because they are “lazy.” They do it because of <strong>how they are trained.</strong></p><p>Large Language Models (LLMs) learn from vast bodies of public code. Because the web has existed for over three decades, public repositories are overwhelmingly saturated with legacy code patterns and outdated stack answers.</p><p>Even if an AI model technically “knows” that a modern browser API exists, it often lacks the density of real-world, modern implementation patterns required to generate clean, production-ready code. Consequently, it defaults to the most common answers in its training data, which usually translate to bloated JavaScript libraries or decade-old CSS workarounds.</p><p><strong>Modern Web Guidance bridges this gap.</strong></p><h3>What is Modern Web Guidance?</h3><p>At its core, <strong>Modern Web Guidance is an expert-curated set of agent skills</strong> that embed modern web platform expertise, best practices, and browser compatibility data directly into your coding agents.</p><p>Instead of letting an AI guess how to solve a problem, this tool inserts targeted, token-efficient guidelines directly into the agent’s context window. It acts as a real-time, expert “advisor” to your IDE assistant, steering it toward solutions that leverage modern, native platform capabilities.</p><p>The guidance covers <strong>over 100 web development use cases</strong> across several core disciplines:</p><ul><li><strong>User Experience &amp; Motion:</strong> Smoothly animating entry and exit states, styling custom scrollbars (scrollbar-color), establishing View Transitions, and creating physics-based easing.</li><li><strong>CSS &amp; Layout:</strong> Transitioning to modern layout features like container queries, CSS subgrid, and utilizing highly responsive color spaces like oklch.</li><li><strong>Forms &amp; UI:</strong> Building accessible forms with native validation styling (like :user-invalid), auto-sizing text input fields (field-sizing: content), and positioning tooltips natively with CSS Anchor Positioning.</li><li><strong>Performance:</strong> Breaking up heavy, long-running JavaScript tasks with scheduler.yield, optimizing Largest Contentful Paint (LCP) image priorities, and deferring heavy rendering with content-visibility.</li><li><strong>Built-in AI &amp; Security:</strong> Accessing on-device client models directly for language translation, detection, or summarization (using Prompt and Summarizer APIs), and implementing secure Passkey authentication.</li></ul><h3>The Secret Weapon: “Baseline” and Intelligent Fallbacks</h3><p>One of the biggest concerns with adopting cutting-edge web features is <strong>browser compatibility.</strong> AI models notoriously make outdated suggestions that fail to take project requirements or real-world browser support into account.</p><p>Modern Web Guidance solves this by integrating with <strong>Baseline</strong>, a standard that categorizes web platform features into three clear support tiers:</p><ol><li><strong>Limited availability:</strong> The feature is not yet fully interoperable across browsers.</li><li><strong>Newly available:</strong> The feature recently became interoperable across all major browsers (within the last 30 months).</li><li><strong>Widely available:</strong> The feature is fully interoperable and has been for at least 30 months.</li></ol><p>By default, if you do not configure a Baseline target for your project, Modern Web Guidance will default to <strong>Baseline Widely available</strong>. However, you can configure your agent to target a different Baseline, such as newly-available, simply by adding a couple of lines of configuration to your project&#39;s AGENTS.md or CLAUDE.md file:</p><pre># AGENTS.md<br>web_platform:<br>  baseline: newly-available</pre><p>When your project targets newer features that are not yet universally supported, the tool instructs your AI agent to write <strong>nuanced, lightweight fallbacks</strong>. Instead of pulling in massive third-party polyfills, it forces the AI to write progressive enhancements, allowing modern browsers to run the native code while conditionally loading tiny, case-specific fallbacks (typically under 50 lines of code) for legacy environments.</p><h3>Getting Started</h3><p>Because Modern Web Guidance is built as an open agent skill, it is compatible with popular AI workflows and coding clients, including <strong>Claude Code, GitHub Copilot, JetBrains WebStorm, Gemini CLI, and Antigravity</strong>.</p><p>To install it and run the interactive setup wizard, you can run the following command directly in your terminal:</p><pre>npx modern-web-guidance@latest install</pre><p>This runs an interactive setup that configures the skills to your preferences.</p><p>If you just want to explore the guides yourself before committing, you can search the library locally using the CLI:</p><pre># Search for how to handle scroll snap highlights<br>npx modern-web-guidance@latest search \&quot;carousel scroll snap\&quot;</pre><p>The CLI will output the relevant use-case IDs, which you can then retrieve directly to view the raw markdown guidelines in your terminal:</p><pre># Retrieve a specific guide<br>npx modern-web-guidance@latest retrieve \&quot;animate-to-from-top-layer\&quot;</pre><h3>Proven Performance</h3><p>Does it actually work? Early automated evaluations using a Playwright-based testing harness show a <strong>37 percentage point improvement</strong> on average in the AI’s adherence to modern best practices. Across popular tools, the unguided-to-guided accuracy uplift was striking, with <strong>Claude Code</strong> seeing a <strong>+35 percentage point</strong> increase (climbing from 52% to 87%) and <strong>Antigravity</strong> improving by <strong>+33 percentage points</strong> (climbing from 54% to 87%).</p><p>Rather than relying on outdated, bloated frameworks, developers can confidently let their agents leverage native platform capabilities, resulting in faster load times, cleaner codebases, and much better user experiences.</p><h3>Get the Latest Updates: Join the Early Preview Program</h3><p>Modern Web Guidance is currently in active development. If you want to stay on the absolute cutting edge, bring the latest standards to your coding workflow, and help shape the project’s direction, you can join the official <a href="http://goo.gle/chrome-ai-dev-preview-join">Early Preview Program</a> directly.</p><p>For those who want to contribute guides or report agent bugs, you can also check out the official open-source <a href="https://github.com/GoogleChrome/modern-web-guidance-src">Modern Web Guidance GitHub repository</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a3e173407d6e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/dogus-teknoloji/fixing-ai-generated-legacy-code-inside-the-modern-web-guidance-initiative-a3e173407d6e">🧭 Fixing AI-Generated Legacy Code: Inside the “Modern Web Guidance” Initiative</a> was originally published in <a href="https://medium.com/dogus-teknoloji">Doğuş Teknoloji</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ WebMCP for Beginners: Connecting AI Agents to Web Interfaces]]></title>
            <link>https://medium.com/dogus-teknoloji/webmcp-for-beginners-connecting-ai-agents-to-web-interfaces-8c0337f7da08?source=rss-ca2093f71a1d------2</link>
            <guid isPermaLink="false">https://medium.com/p/8c0337f7da08</guid>
            <category><![CDATA[web-mcp]]></category>
            <category><![CDATA[ai-web]]></category>
            <category><![CDATA[chrome]]></category>
            <category><![CDATA[javascript]]></category>
            <category><![CDATA[html]]></category>
            <dc:creator><![CDATA[Ege Kaan Işık]]></dc:creator>
            <pubDate>Thu, 09 Jul 2026 12:50:13 GMT</pubDate>
            <atom:updated>2026-07-16T09:22:33.152Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*miCh-R1VEfgy1mPAglV_aA.jpeg" /></figure><p>For years, browser automation has been a complex challenge. When developers wanted an AI agent to interact with a website, the agent had to rely on “screen-scraping” or “actuation” — painstakingly simulating mouse clicks and keyboard inputs. If a button shifted slightly to the left, or a class name changed during a routine update, the entire integration shattered.</p><p>With <strong>WebMCP (Web Model Context Protocol)</strong>, a proposed web standard, web development is transitioning away from fragile screen-scraping toward a robust, structured, and secure contract between websites and browser-based AI agents. Instead of requiring an agent to guess what a button does, websites can now explicitly declare their capabilities to the browser’s built-in AI.</p><p>Whether building shopping assistants, automated form-fillers, or complex in-page workflow automations, WebMCP provides the missing architectural bridge. This guide breaks down how the protocol works, how it compares to the server-side Model Context Protocol (MCP), and how to implement it today.</p><h3>WebMCP vs. MCP: Partners, Not Replacements</h3><p>As AI agents become a native part of the browsing experience, a common question arises: <em>“Is WebMCP replacing the Model Context Protocol (MCP)?”</em></p><p>The answer is: <strong>No. They are partners, not competitors.</strong> They solve fundamentally different parts of the agentic integration puzzle:</p><ul><li><strong>MCP (Model Context Protocol)</strong> is a persistent, backend-focused standard. It runs on a server or as a local daemon to connect agents to external databases, enterprise files, and background workflows. It is platform-agnostic and runs 24/7.</li><li><strong>WebMCP</strong> is a frontend-focused, client-side browser standard. It is entirely ephemeral — active only while the user has the website open in a browser tab. It is DOM-aware, secure, and has instant, zero-latency access to the active page’s state, local cookies, and user session.</li></ul><p>Think of it like this: <strong>MCP is a company’s backend customer service call center. WebMCP is the helpful, in-store expert standing right next to the user in your shop.</strong></p><p>By combining both, developers can manage core business logic via MCP servers globally, while using WebMCP to give the browser agent instant, high-fidelity control over the active UI in the user’s current tab.</p><h3>The Dual-API System: Imperative vs. Declarative</h3><p>WebMCP offers two distinct paradigms to expose tools to browser agents. Developers can register programmatic functions using standard JavaScript (the <strong>Imperative API</strong>) or enhance existing HTML markup (the <strong>Declarative API</strong>).</p><h3>1. The Imperative API (JavaScript)</h3><p>The Imperative API allows the registration of complex, programmatic tools directly within client-side JavaScript code.</p><blockquote><strong><em>Important Deprecation Note:</em></strong><em> If you are studying early WebMCP documentations, note that </em><em>navigator.modelContext has been deprecated in Chrome 150 in favor of </em><em>document.modelContext .</em></blockquote><p>Below is an example of registering a tool dynamically using JavaScript:</p><pre>// Define a tool to add an item to a todo list<br>const addTodoTool = {<br>  name: &quot;addTodo&quot;,<br>  description: &quot;Add a new item to the user&#39;s active todo list&quot;,<br>  inputSchema: {<br>    type: &quot;object&quot;,<br>    properties: {<br>      text: { <br>        type: &quot;string&quot;, <br>        description: &quot;The plain-text task content to add&quot; <br>      }<br>    },<br>    required: [&quot;text&quot;]<br>  },<br>  execute: ({ text }) =&gt; {<br>    // Handle the task addition and UI updates<br>    console.log(`Successfully added: ${text}`);<br>    return `Added todo: ${text}`;<br>  },<br>  annotations: { <br>    readOnlyHint: false, <br>    untrustedContentHint: true <br>  }<br>};<br><br>// Register the tool<br>const controller = new AbortController();<br>document.modelContext.registerTool(addTodoTool, { signal: controller.signal });<br><br>// If the user navigates away or cleanup is required:<br>// controller.abort();</pre><p>With this registration completed, any compatible browser agent can instantly discover the tool, parse its parameters via the generated JSON Schema, and invoke it programmatically.</p><h3>2. The Declarative API (HTML Forms)</h3><p>For simpler forms of integration, standard HTML forms can be transformed into AI-ready tools by adding specific annotations. The browser automatically parses the form elements, generates the computed JSON Schema, and exposes the tool to the active agent context.</p><p>Consider the following declarative form designed to let an agent submit a customer support request:</p><pre>&lt;form <br>  toolname=&quot;submit_support_ticket&quot; <br>  tooldescription=&quot;Submit a support ticket for technical assistance&quot; <br>  toolautosubmit<br>  action=&quot;/api/support&quot;<br>&gt;<br>  &lt;label for=&quot;category&quot;&gt;Support Category&lt;/label&gt;<br>  &lt;select <br>    id=&quot;category&quot; <br>    name=&quot;category&quot; <br>    required <br>    toolparamdescription=&quot;The general category of the technical issue (e.g., billing, login, hardware)&quot;<br>  &gt;<br>    &lt;option value=&quot;billing&quot;&gt;Billing Inquiry&lt;/option&gt;<br>    &lt;option value=&quot;login&quot;&gt;Account Access&lt;/option&gt;<br>    &lt;option value=&quot;bug&quot;&gt;Technical Bug&lt;/option&gt;<br>  &lt;/select&gt;<br>  <br>  &lt;label for=&quot;details&quot;&gt;Problem Details&lt;/label&gt;<br>  &lt;textarea id=&quot;details&quot; name=&quot;details&quot; required&gt;&lt;/textarea&gt;<br>  <br>  &lt;button type=&quot;submit&quot;&gt;Submit Ticket&lt;/button&gt;<br>&lt;/form&gt;</pre><p>When an agent invokes a declarative tool:</p><ol><li>The browser automatically focuses the form and pre-fills the fields.</li><li>Because toolautosubmit is present, it will automatically submit and navigate without forcing the user to manually click &quot;Submit&quot;.</li><li>For custom client-side validation, the SubmitEvent now exposes agentInvoked (a boolean indicating if an AI triggered the form) and respondWith(Promise) to intercept the submission and serialize the response back to the model.</li></ol><pre>document.querySelector(&quot;form&quot;).addEventListener(&quot;submit&quot;, (e) =&gt; {<br>  e.preventDefault();<br>  <br>  if (e.agentInvoked) {<br>    // Intercept agent action, run validation, and return JSON directly to the model<br>    e.respondWith(Promise.resolve(&quot;Ticket successfully logged via WebMCP!&quot;));<br>  }<br>});</pre><h3>Security Best Practices: Guarding Against Injection</h3><p>Exposing application state and form submissions to an AI agent provides immense capabilities, but it requires rigorous security guardrails. Because LLMs process instructions and external data within a single sequence of tokens, they are susceptible to <strong>indirect prompt injection</strong> (where malicious instructions are embedded inside third-party web content, user-generated reviews, or emails).</p><p>To protect both the application and the user, developers should implement several core guardrails:</p><h3>Use Annotation Hints</h3><p>Always apply security hints to guide the browser agent’s interaction:</p><ul><li>untrustedContentHint: true: Add this to any tool that processes or returns user-generated content (UGC) or unverified external data. This signals to the agent that the payload requires extreme scrutiny.</li><li>readOnlyHint: true: Set this on tools that retrieve data without changing state. This allows the agent to make smart decisions about when to ask for user confirmation.</li></ul><h3>Observe Origin Isolation &amp; Permissions</h3><p>WebMCP only operates in secure, origin-isolated documents (ensured by headers like Origin-Agent-Cluster). Furthermore, WebMCP tools are gated by a default tools Permissions Policy. If you need to expose tools inside cross-origin iframes, you must explicitly delegate access using allow=&quot;tools&quot;:</p><pre>&lt;iframe src=&quot;https://trustedpartner.com&quot; allow=&quot;tools&quot;&gt;&lt;/iframe&gt;</pre><h3>Stick to Character Budgets</h3><p>To keep agent reasoning accurate and avoid hitting context limits, follow these recommended budgets:</p><ul><li><strong>Tool Description</strong>: Max 500 characters</li><li><strong>Parameter Description</strong>: Max 150 characters</li><li><strong>Names (Tools &amp; Parameters)</strong>: Max 30 characters</li><li><strong>Tool Output Payload</strong>: Max 1.5K characters per individual output</li></ul><h3>How to Get Started Today</h3><p>WebMCP is currently available as a proposed standard under active developer preview. Developers can begin experimenting with local integration by following these steps:</p><ol><li><strong>Enable the Chrome Flag</strong>: Open Chrome (version 146 or higher) and navigate to chrome://flags/#enable-webmcp-testing. Set the flag to <strong>Enabled</strong> and relaunch your browser.</li><li><strong>Install the Inspector</strong>: Download the official <a href="https://chromewebstore.google.com/detail/model-context-tool-inspec/gbpdfapgefenggkahomfgkhfehlcenpd">Model Context Tool Inspector Extension</a>. This extension provides an invaluable developer UI to monitor registered tools, execute them manually with raw arguments, and run simulated agent prompts using your own Gemini API keys.</li><li><strong>Explore the Codebases</strong>: Dive into experimental command-line tooling in the evals-cli folder of the <a href="https://github.com/GoogleChromeLabs/webmcp-tools/tree/main/evals-cli">official WebMCP Tools GitHub Repository</a>.</li></ol><h3>Join the Community!</h3><p>As an evolving proposal, WebMCP benefits directly from developer feedback to shape its ergonomics, security infrastructure, and tooling ecosystem.</p><p>To stay updated on the latest API specifications, participate in developer discussions, and join the dedicated mailing list, developers are encouraged to sign up for the <strong>Early Preview Program (EPP)</strong>:</p><p>👉 <a href="http://goo.gle/chrome-ai-dev-preview-join"><strong>Join the WebMCP Early Preview Program (EPP)</strong></a></p><p>By adopting WebMCP today, developers can ensure their applications are fully prepared to participate in the emerging browser agent ecosystem safely and reliably.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8c0337f7da08" width="1" height="1" alt=""><hr><p><a href="https://medium.com/dogus-teknoloji/webmcp-for-beginners-connecting-ai-agents-to-web-interfaces-8c0337f7da08">🔌 WebMCP for Beginners: Connecting AI Agents to Web Interfaces</a> was originally published in <a href="https://medium.com/dogus-teknoloji">Doğuş Teknoloji</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ Chrome Yerleşik AI API’leri: Gemini Nano ile Cihaz Üzerinde Yapay Zekayı Güçlendirmek]]></title>
            <link>https://medium.com/dogus-teknoloji/chrome-yerle%C5%9Fik-ai-apileri-gemini-nano-ile-cihaz-%C3%BCzerinde-yapay-zekay%C4%B1-g%C3%BC%C3%A7lendirmek-5772e0dcff3f?source=rss-ca2093f71a1d------2</link>
            <guid isPermaLink="false">https://medium.com/p/5772e0dcff3f</guid>
            <category><![CDATA[gemini]]></category>
            <category><![CDATA[ai-web]]></category>
            <category><![CDATA[chrome]]></category>
            <category><![CDATA[yapay-zeka]]></category>
            <category><![CDATA[javascript]]></category>
            <dc:creator><![CDATA[Ege Kaan Işık]]></dc:creator>
            <pubDate>Wed, 17 Dec 2025 09:19:30 GMT</pubDate>
            <atom:updated>2025-12-17T17:29:39.772Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2II9n7kv5KcBW6LmRig_sw.png" /></figure><p>Web geliştirmede yapay zeka uzun süre aynı mimariyle ele alındı: Bir backend, bir model, bir API endpoint’i ve istemciden yapılan çağrılar.</p><p>Bugün ise web tarafında <strong>daha katmanlı ve olgun bir yapay zeka ekosistemi</strong> oluşmuş durumda. Bu ekosistemin genel adı: <strong>Web AI</strong>.</p><p>Chrome’un tarayıcıyla birlikte sunduğu <strong>yerleşik AI API’leri</strong>, Web AI’ın geliştiriciye en yakın katmanını temsil ediyor: <strong>hazır, opinionated ve doğrudan kullanılabilir yapay zeka yetenekleri.</strong></p><h3>🌐 Web AI Nedir?</h3><p><strong>Web AI</strong>, yapay zeka yeteneklerinin web platformu üzerinde, tarayıcı ortamında veya tarayıcıya çok yakın şekilde çalıştırılmasını hedefleyen yaklaşımların genel adıdır.</p><p>Temel fikir şudur:<br>Yapay zeka, web için dışarıdan bağlanan bir servis olmak zorunda değildir; <strong>web platformunun doğal bir parçası olabilir.</strong></p><p>Bu ekosistemde farklı seviyelerde çözümler bulunur:</p><ul><li>🔧 <strong>Düşük seviye yaklaşımlar<br></strong><em>MediaPipe</em>: Görüntü, ses ve multimodal görevler<br><em>LiteRT</em>: Modeli sen getir, inference sürecini sen yönet</li><li>⚡ <strong>Yüksek seviye yaklaşımlar<br></strong>Tarayıcı tarafından hazır sunulan, doğrudan kullanılabilir AI yetenekleri</li></ul><p>Yerleşik AI API’leri bu ikinci grupta yer alır.</p><h3>🎯 Task Odaklı Yerleşik AI Yaklaşımı</h3><p>Yerleşik AI API’leri, Web AI ekosistemi içinde <strong>task odaklı</strong> bir katman olarak konumlanır.</p><p>Bu API’ler sana şunu dedirtir: “Modeli bilmeme gerek yok, <strong>ben ne yapmak istiyorum?</strong>”</p><p>Amaç:</p><ul><li>🌍 Metni çevirmek</li><li>✂️ Özetlemek</li><li>✍️ Yeniden yazmak</li><li>✅ Düzeltmek</li></ul><p>Yani AI’yi ayrı bir ürün gibi değil, <strong>günlük web geliştirme akışının doğal bir parçası</strong> olarak kullanmak.</p><h3>⚙️ Başlamadan Önce</h3><p>Bu API’ler tarayıcıya gömülü modeller kullandığı için klasik Web API’lerinden biraz farklıdır.</p><p>Kısaca:</p><ul><li>🖥️ Güncel <strong>masaüstü Chrome</strong> gerekir</li><li>🚩 Bazı özellikler <strong>Chrome flags</strong> ile açılabilir</li><li>🧪 Bazı API’ler <strong>erken erişim (EPP)</strong> kapsamında olabilir</li><li>📱 Mobil destek sınırlıdır</li></ul><p>👉 Güncel gereksinimler, flag’ler ve EPP durumu için <a href="https://developer.chrome.com/docs/ai/built-in-apis"><strong>resmi Chrome dokümantasyonuna</strong></a><strong> bakmak gerekir</strong>.</p><h3>🧩 Task Odaklı Yerleşik AI API’leri</h3><h3>🌍 Translator API</h3><p><strong>API adı:</strong> Translator</p><p>Kullanıcıdan gelen içeriği başka bir dile çevirmek çoğu zaman backend’e taşınan bir iştir. Translator API bu işi doğrudan tarayıcıya alır.</p><pre>const availability = await Translator.availability({<br>  sourceLanguage: &#39;en&#39;,<br>  targetLanguage: &#39;tr&#39;<br>});<br><br>if (availability !== &#39;unavailable&#39;) {<br>  const translator = await Translator.create({<br>    sourceLanguage: &#39;en&#39;,<br>    targetLanguage: &#39;tr&#39;<br>  });<br>  const result = await translator.translate(<br>    &#39;Welcome to Web AI&#39;<br>  );<br>  console.log(result);<br>  translator.destroy();<br>}</pre><h3>🗣️ Language Detector API</h3><p><strong>API adı:</strong> LanguageDetector</p><p>Bir metnin hangi dilde yazıldığını otomatik olarak tespit eder.</p><pre>const availability = await LanguageDetector.availability();<br><br>if (availability !== &#39;unavailable&#39;) {<br>  const detector = await LanguageDetector.create();<br>  const results = await detector.detect(<br>    &#39;Merhaba dünya&#39;<br>  );<br>  console.log(results);<br>  detector.destroy();<br>}</pre><h3>📄 Summarizer API</h3><p><strong>API adı:</strong> Summarizer</p><p>Uzun içerikleri hızlıca tüketilebilir hale getirir.</p><pre>const availability = await Summarizer.availability();<br><br>if (availability !== &#39;unavailable&#39;) {<br>  const summarizer = await Summarizer.create({<br>    type: &#39;key-points&#39;<br>  });<br>  const summary = await summarizer.summarize(<br>    &#39;Web AI enables AI directly in the browser...&#39;<br>  );<br>  console.log(summary);<br>  summarizer.destroy();<br>}</pre><h3>✍️ Writer API</h3><p><strong>API adı:</strong> Writer</p><p>Kullanıcıya içerik üretiminde başlangıç noktası sunar.</p><pre>const availability = await Writer.availability();<br><br>if (availability !== &#39;unavailable&#39;) {<br>  const writer = await Writer.create();<br>  const text = await writer.write(<br>    &#39;Introduce Web AI to frontend developers&#39;<br>  );<br>  console.log(text);<br>  writer.destroy();<br>}</pre><h3>🔁 Rewriter API</h3><p><strong>API adı:</strong> Rewriter</p><p>Mevcut metni anlamı korunacak şekilde dönüştürür.</p><pre>const availability = await Rewriter.availability();<br><br>if (availability !== &#39;unavailable&#39;) {<br>  const rewriter = await Rewriter.create();<br>  const rewritten = await rewriter.rewrite(<br>    &#39;This is kinda cool&#39;,<br>    { tone: &#39;more-formal&#39; }<br>  );<br>  console.log(rewritten);<br>  rewriter.destroy();<br>}</pre><h3>🪄 Proofreader API</h3><p><strong>API adı:</strong> Proofreader</p><p>Yazım ve dilbilgisi hatalarını düzeltir.</p><pre>const availability = await Proofreader.availability();<br><br>if (availability !== &#39;unavailable&#39;) {<br>  const proofreader = await Proofreader.create();<br>  const fixedText = await proofreader.proofread(<br>    &#39;This text have a mistake&#39;<br>  );<br>  console.log(fixedText);<br>  proofreader.destroy();<br>}</pre><h3>🧠 Prompt API: Task’lerin Ötesinde, Doğrudan Model Erişimi</h3><p><strong>API adı:</strong> LanguageModel</p><p>Prompt API, belirli bir task yerine <strong>doğrudan Gemini Nano modeline erişim</strong> sağlar.</p><pre>const availability = await LanguageModel.availability();<br><br>if (availability !== &#39;unavailable&#39;) {<br>  const model = await LanguageModel.create();<br>  const output = await model.prompt(<br>    &#39;Explain Web AI in one sentence&#39;<br>  );<br>  console.log(output);<br>  model.destroy();<br>}</pre><h3>🚀 Kapanış</h3><p>Web AI tek bir teknoloji değil, <strong>katmanlı bir ekosistem</strong>. Yerleşik AI API’leri ise bu ekosistemin geliştiriciye en yakın katmanı.</p><p>Tarayıcı artık sadece render etmiyor; metni anlıyor, dönüştürüyor ve iyileştiriyor. Prompt API ile birlikte, <strong>gerekirse doğrudan modelle de konuşabiliyor.</strong></p><p><strong>Web artık düşünüyor — ve bunu tarayıcının içinde yapıyor.</strong></p><h3>🔗 Kaynaklar</h3><ul><li><strong>Built-in AI APIs — Chrome Documentation<br></strong><a href="https://developer.chrome.com/docs/ai/built-in-apis?utm_source=chatgpt.com">https://developer.chrome.com/docs/ai/built-in-apis</a></li><li><strong>Early Preview Program (EPP)<br></strong><a href="https://developer.chrome.com/docs/ai/join-epp">https://developer.chrome.com/docs/ai/join-epp</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5772e0dcff3f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/dogus-teknoloji/chrome-yerle%C5%9Fik-ai-apileri-gemini-nano-ile-cihaz-%C3%BCzerinde-yapay-zekay%C4%B1-g%C3%BC%C3%A7lendirmek-5772e0dcff3f">🤖 Chrome Yerleşik AI API’leri: Gemini Nano ile Cihaz Üzerinde Yapay Zekayı Güçlendirmek</a> was originally published in <a href="https://medium.com/dogus-teknoloji">Doğuş Teknoloji</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[️ Siber Savunmada Rickroll Stratejisi: Bir URL Kısaltma Servisinin Güvenlik Macerası]]></title>
            <link>https://medium.com/dogus-teknoloji/%EF%B8%8F-siber-savunmada-rickroll-stratejisi-bir-url-k%C4%B1saltma-servisinin-g%C3%BCvenlik-maceras%C4%B1-245f1643be94?source=rss-ca2093f71a1d------2</link>
            <guid isPermaLink="false">https://medium.com/p/245f1643be94</guid>
            <category><![CDATA[yazilim]]></category>
            <category><![CDATA[cloudflare]]></category>
            <category><![CDATA[web]]></category>
            <category><![CDATA[security]]></category>
            <category><![CDATA[siber-güvenlik]]></category>
            <dc:creator><![CDATA[Ege Kaan Işık]]></dc:creator>
            <pubDate>Tue, 17 Jun 2025 11:23:37 GMT</pubDate>
            <atom:updated>2025-06-17T12:25:33.306Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ErpN2qibt0oOkT_hMskjsQ.png" /></figure><p>URL kısaltma servisleri, internetin olmazsa olmazlarından. Peki ya bu servisler, siber saldırganların hedefi haline gelirse? Bu yazıda, <strong>GDG İstanbul</strong> için geliştirdiğim bir URL kısaltma servisinin, beklenmedik siber saldırılarla nasıl başa çıktığını ve bu süreçte <strong>Rick Astley’den</strong> ilham alan yaratıcı savunma stratejilerini keşfedeceksiniz. Kendi sistemim üzerindeki bu macera, siber güvenlikte aktif mücadele ile statik koruma arasındaki ince çizgiyi anlamamı sağladı.</p><h3>🚀 Projenin Doğuşu ve Beklenmedik “Misafirler”</h3><p>GDG İstanbul etkinlikleri için pratik ve akılda kalıcı URL’lere olan ihtiyacımızı karşılamak amacıyla, <strong>go-gdg.ist</strong> alan adı üzerinde özel bir URL kısaltma servisi geliştirdim. Bu projede <strong>Express.js</strong> ve <strong>Strapi CMS</strong> teknolojilerini kullanarak, gelen istekleri ilgili bağlantılara veya dosyalara yönlendiren, bulunamayanlar için ise 404 sayfası sunan basit ama etkili bir yapı kurdum. Hizmetin performansını ve kullanım istatistiklerini takip etmek amacıyla tüm yönlendirmeleri <strong>Google Analytics</strong> üzerinden detaylı bir şekilde kaydettim.</p><p>Sistem beklenen şekilde işlerken, bir gün <strong>Analytics</strong> verilerimi incelerken <strong>“not_found”</strong> etkinliklerinde belirgin bir artış gözlemledim. Kısa bir süre içinde yaklaşık <strong>2000&#39;e yakın</strong> başarısız istekle karşılaşmıştım. İsteklerin hedeflediği endpoint’lere baktığımda ise durum netleşti: /wp-admin/setup-config.php, /wordpress/wp-admin/setup-config.php, /.git/config gibi tipik <strong>zaafiyet tarama</strong> veya <strong>saldırı denemesi</strong> yollarına erişilmeye çalışıldığı anlaşılıyordu.</p><h3>🎣 Honeypot Stratejisi ve “Rick Astley” Sürprizi</h3><p>Bu durumu tespit ettikten sonra, hem bir güvenlik önlemi almak hem de duruma yaratıcı bir yaklaşım katmak istedim. Saldırganların bilinen zafiyetleri hedef almasından yola çıkarak, bu “istenmeyen” endpoint’leri <strong>Rick Astley’nin “Never Gonna Give You Up” YouTube linkine yönlendirmeye</strong> karar verdim. Beklentim, bu şaşırtıcı yönlendirme sonrası denemelerin azalmasıydı. Ancak, tam tersi bir etkiyle karşılaştım; <strong>istekler daha da arttı!</strong> Görünüşe göre siber saldırganlar, hedeflerinde bir sürprizle karşılaşınca daha da motive olabiliyordu.</p><p>Bu gelişme üzerine stratejimi bir adım ileri taşıdım. <strong>ChatGPT</strong>’den sıkça hedeflenen zafiyet tarama endpoint’lerinin bir listesini alarak, <strong>Strapi</strong> üzerinde yeni bir <strong>collection</strong> oluşturdum. Bu listedeki tüm endpoint’leri yine <strong>“Never Gonna Give You Up”a</strong> yönlendirirken, aynı zamanda bu <strong>honeypot’a</strong> düşen her isteğin <strong>IP adresini</strong> de detaylı bir şekilde kaydetmeye başladım.</p><p>Elde ettiğim IP adreslerini kullanarak, <strong>Cloudflare</strong> üzerinde bir <strong>“Deny” kuralı</strong> oluşturdum ve ilk etapta belirlediğim bazı IP’leri engelledim. Bu adımla sürecin sona ereceğini umarken, farklı IP’lerden honeypot’a yönelik isteklerin devam ettiğini gördüm. Bu durum, pasif savunmanın yeterli olmadığını ve daha <strong>aktif bir mücadele</strong> gerektirdiğini ortaya koydu.</p><h3>⚔️ Aktif Savunma ve IP Quality Score Entegrasyonu</h3><p>Aktif bir savunma mekanizması kurmak adına, honeypot’lara gelen her isteği gerçek zamanlı olarak taramaya başladım. Gelen IP’leri <strong>IP Quality Score (IPQS) API’si</strong> ile analiz ettim:</p><ul><li>Eğer bir IP’nin <strong>fraud skoru 85&#39;in üzerindeyse</strong>, bu durum yüksek riskli bir aktiviteye işaret ediyordu. Bu IP’yi hemen engellemek için <strong>Cloudflare API</strong>’sini kullandım. Mevcut “Deny” kuralımın ifadesini dinamik olarak güncelleyerek, yeni IP’leri bu kurala ekledim. (Cloudflare’in ücretsiz katmanının <strong>5 kural ile sınırlı olması</strong>, yeni kurallar oluşturmak yerine mevcut kuralı modifiye etmeyi tercih etmeme neden oldu.)</li><li>Eğer <strong>fraud skoru 85&#39;in altındaysa</strong>, bu IP’yi izlemeye aldım. Aynı IP’den <strong>5 kez honeypot’a istek gelmesi</strong> durumunda, bunu da kötü amaçlı bir deneme olarak kabul edip aynı yöntemle Cloudflare üzerinden engelledim.</li></ul><p>Bu proaktif sistem kısa sürede başarılı bir şekilde işlemeye başladı. Ancak bu mücadelenin dinamik bir süreç olduğunu, saldırganların da sürekli yeni yöntemler denediğini tecrübe ettim.</p><h3>🧱 Sürekli Bir Yarış ve Nihai Çözüm: Statik Savunma</h3><p>Bir süre sonra, saldırganların <strong>benim honeypot olarak belirlemediğim endpoint’leri</strong> denemeye başladığını fark ettim. Bu yeni hedefler, tanımlı izleme mekanizmalarımın dışında kaldığı için IPQS veya honeypot isabet kontrollerine dahil olmuyor, dolayısıyla istekler geçerli kabul ediliyordu.</p><p>Tüm gelen istekleri IPQS API’si ile kontrol etme fikrini düşündüm, ancak bu sefer de <strong>API limitlerine takılma riski</strong> ortaya çıktı. İşte tam bu noktada, önemli bir çıkarıma ulaştım: <strong>Saldırganlarla ne kadar aktif mücadele edersem edeyim, daima bir adım geride kalma potansiyelim vardı.</strong> Bu sürekli takip ve yanıtlama döngüsü, sürdürülebilir bir çözüm değildi.</p><p>Bu farkındalık beni daha köklü bir yaklaşıma yöneltti: <strong>Pasif ve güçlü bir savunma altyapısı kurmak.</strong> Eğer savunma mekanizmalarım o kadar sağlam olursa ki, saldırganların ne kadar denese de içeri sızamayacağını bilirsem, hem operasyonel yüküm azalacak hem de sistemin güvenliği garanti altına alınacaktı.</p><p>Şu an itibarıyla, URL kısaltma servisimin altyapısı sadece <strong>GET metodlarını</strong> kabul edecek şekilde yapılandırılmıştır. Ayrıca <strong>Cloudflare’in gelişmiş bot savaş modu</strong> ve <strong>kapsamlı DNS güvenlik önlemleri</strong> aktif durumdadır. Bu entegre savunma katmanları, sisteme yönelik riskleri minimize etmektedir.</p><h3>💡 Nihai Ders: Proaktif Savunmadan Statik Korumaya Geçiş</h3><p>Bu tecrübe bana gösterdi ki, siber güvenlikte <strong>her saldırıya tek tek yanıt vermeye çalışmak, verimsiz ve yorucu bir çabadır.</strong> Bunun yerine, <strong>kalıcı ve sağlam güvenlik mimarileri inşa etmek, pasif savunmayı güçlendirmek</strong>, uzun vadede çok daha etkili ve sürdürülebilir bir çözümdür.</p><p>Sonuç olarak, go-gdg.ist şu an güvende ve Rick Astley, artık sadece ben istediğim zaman çalıyor. 😂</p><p>Siz de benzer siber güvenlik deneyimleri yaşadınız mı? Yaşadıysanız, bu konudaki stratejilerinizi ve çıkarımlarınızı yorumlarda paylaşmaktan çekinmeyin!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=245f1643be94" width="1" height="1" alt=""><hr><p><a href="https://medium.com/dogus-teknoloji/%EF%B8%8F-siber-savunmada-rickroll-stratejisi-bir-url-k%C4%B1saltma-servisinin-g%C3%BCvenlik-maceras%C4%B1-245f1643be94">🛡️ Siber Savunmada Rickroll Stratejisi: Bir URL Kısaltma Servisinin Güvenlik Macerası</a> was originally published in <a href="https://medium.com/dogus-teknoloji">Doğuş Teknoloji</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>