<?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 Budhdi Sharma on Medium]]></title>
        <description><![CDATA[Stories by Budhdi Sharma on Medium]]></description>
        <link>https://medium.com/@budhdisharma?source=rss-779303a3099------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*pWzPN9z_gwfbmbUhqhIcfA.jpeg</url>
            <title>Stories by Budhdi Sharma on Medium</title>
            <link>https://medium.com/@budhdisharma?source=rss-779303a3099------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 08:50:14 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@budhdisharma/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[What Is JEV? Inside the AI Model Built to Replace LLMs for Fast Decisions]]></title>
            <link>https://medium.com/android-knowledge-store/what-is-jev-inside-the-ai-model-built-to-replace-llms-for-fast-decisions-5f6645c01df5?source=rss-779303a3099------2</link>
            <guid isPermaLink="false">https://medium.com/p/5f6645c01df5</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[software-architecture]]></category>
            <dc:creator><![CDATA[Budhdi Sharma]]></dc:creator>
            <pubDate>Thu, 24 Sep 2026 17:40:44 GMT</pubDate>
            <atom:updated>2026-09-24T17:40:44.365Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xiBUwCvb5q05AcYZq82tJQ.png" /><figcaption>JEV Model</figcaption></figure><p>It was close to midnight, and I was doing the thing I always do before I actually sleep, scrolling through half-finished tabs, three of them YouTube, two of them GitHub. Every second video on my feed had the same word stamped across the thumbnail: <strong>JEV</strong>. Not GPT, not Claude, not Gemini. A name I hadn’t heard a month ago, suddenly everywhere.</p><p>My first instinct, honestly, was to ignore it. I’ve been building on embedded systems long enough — almost a decade in embedded and infotainment stacks, and the last few years knee-deep in ML architecture — to know that most “the next big thing” claims quietly die within two weeks. Noise-to-signal in this field is brutal.</p><p>But this one felt different. Not because of the hype. Because of one line I kept running into, over and over, from the person who built it:</p><blockquote>Models have been superhuman at chat for years. So where is all the automation?</blockquote><p>That question sat with me. Because I’ve felt that exact frustration myself — building agentic pipelines that work beautifully in a demo and then quietly fall apart the moment they hit production traffic. Slow. Expensive. Occasionally making things up when I least expected it. So I stopped scrolling, opened a blank editor, and spent the next two days actually pulling JEV apart, reading everything public about it, running my own tests, and sketching its architecture on paper the way I do with every system I want to actually understand.</p><p>This article is that notebook, cleaned up.</p><h3><strong>First, the uncomfortable question JEV is built to answer</strong></h3><p>Every LLM you and I use daily GPT, Claude, Gemini, Llama, is what’s called <strong>auto-regressive</strong>. It writes one word, looks at what it just wrote, then writes the next word, and so on until the sentence is done. That’s <em>why</em> it can hold a conversation, write an essay, or explain a bug to you in plain English.</p><p>But here’s the thing nobody says out loud enough: <strong>most of what software actually needs from AI isn’t conversation. It’s a decision.</strong></p><ul><li>Which support queue does this email belong to?</li><li>Is this comment toxic or not?</li><li>Should this agent call the search tool, or answer from memory?</li><li>Is this transaction fraud, yes or no?</li></ul><p>None of that needs an essay. None of it needs “thinking.” It needs an answer, in milliseconds, for a fraction of a rupee. And that gap between what LLMs are <em>great</em> at and what most production software actually <em>needs</em> — is exactly the gap JEV was built to fill.</p><p>Its own definition, and I think it’s a fair one: <strong>JEV is an AI model built to make fast, structured decisions that software can use directly.</strong> Not a chatbot. A decision engine.</p><h3>Borrowing a page from Kahneman</h3><p>If you’ve read <em>Thinking, Fast and Slow</em>, this next part will click immediately.</p><p>Daniel Kahneman split human thinking into two systems. <strong>System 1</strong> is instinctive — you’re on a bike, a car cuts in front of you, you swerve without “deciding” to. <strong>System 2</strong> is deliberate — planning your next five years takes sitting down and actually reasoning it out.</p><p>Today’s LLMs are System 2 machines. They reason, they plan, they think step by step before answering — which is exactly what makes them slow and expensive for the kind of split-second call a piece of software makes a thousand times a minute.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2ajY3yII67mUU2cP8kHVmA.png" /></figure><p>JEV’s whole pitch is that we’ve been using a System-2 hammer on System-1 nails. Routing an email doesn’t need reasoning. It needs an instinct.</p><h3>The numbers that made me actually pay attention</h3><p>I didn’t take the speed and cost claims at face value — I ran the comparison myself, the same way it was demoed: one query, one context, four options, fired at both a GPT-5-class model and at JEV, side by side.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UzKq5UuieOLNILUyZ6UofQ.png" /><figcaption>comparision</figcaption></figure><p>Same question. Same four options. Same correct answer. One model took over six and a half seconds to get there. The other did it in under half a second — roughly <strong>13x faster</strong>, and that gap held up across repeated runs.</p><p>Then I did the maths every engineer eventually has to do — the cost math, not the demo math. Take a support inbox running 10,000 emails a day through a classifier, every single day, for a year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7uDwqdpZ_OIKX5dQcHChMg.png" /><figcaption>cost</figcaption></figure><p>Roughly <strong>₹11 lakh a year</strong> on one side. Roughly <strong>₹14,000</strong> on the other. That’s not a rounding difference — that’s the difference between a line item your CFO questions and one nobody even notices.</p><h3>What’s actually happening inside it</h3><p>This is the part I found genuinely interesting to work through, and it’s also the part where I have to be honest with you: <strong>TypeSafe, the company behind JEV, hasn’t published a paper, a dataset, or an architecture diagram.</strong> No transparency at all, just a website, some claims, and an API. So everything from here is reconstructed reasoning, not confirmed fact. I sat with a pen, the handful of statements they <em>have</em> made public, and a benchmark result, and worked backward.</p><p>Here’s what we know for certain, in their own words: it’s transformer-based, it has broad world knowledge, it’s built for “System 1” tasks, it’s trained on synthetic data, it’s <strong>non-auto-regressive</strong>, it gives schema-constrained output, it does parallel sampling of questions, and it produces a calibrated confidence score.</p><p>The single clue that unlocked the rest for me was a third-party benchmark score: JEV scores around <strong>84.6% on MMLU</strong> — a knowledge benchmark. That kind of world-knowledge score only comes from large-scale pretraining, and large-scale pretraining, practically speaking, only really works on decoder architectures. So despite behaving like a classifier — despite feeling like BERT in spirit — this has to be a <strong>decoder</strong>, the same lineage GPT and Claude come from. It’s just a decoder that’s been stopped from doing the one thing decoders are famous for.</p><p>Every transformer-based LLM runs inference in two stages: <strong>prefill</strong>, where it reads and “understands” your input by turning it into key/value vectors, and <strong>decode</strong>, where it generates the answer one token at a time using something called an LM head — a layer that scores every single word in the vocabulary and picks the most likely next one.</p><p>JEV keeps the prefill stage exactly as-is. It’s the decode stage they tore out and replaced.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hf9BBCRDtDnf0vj1uAToiw.png" /><figcaption>Architecture</figcaption></figure><p>Instead of scoring 50,000+ possible words, JEV’s <strong>answer head</strong> only ever scores the handful of options <em>you</em> hand it, runs a softmax across just those, and returns a probability for each — in one shot, no token-by-token loop. That single swap explains almost every headline claim: it’s faster because it’s not generating text word by word. It’s cheap because there’s no long output to pay for. And it literally cannot hallucinate an answer, because it is mathematically incapable of returning anything outside the options you gave it.</p><p>My best guess on how it was trained, reasoning from the fact that they explicitly say they used <em>only</em> synthetic data: they most likely took an existing open-source pretrained model, kept its prefill/understanding layers intact, bolted on this new answer head, and trained that new head on synthetic question-option-answer triples. A second training pass — using a technique they call <strong>RLCD</strong> (reinforcement learning for calibrated decisions, in the spirit of the RLHF used to align chat models) — is what teaches it to report a confidence score it can actually be trusted on, rather than the kind of confidently-wrong number LLMs are known for.</p><h3>Where the parallel sampling trick comes in</h3><p>The part of the demo that impressed me most wasn’t the speed on a single question — it was what happens when you ask several at once.</p><p>In one run, a single customer email was sent to JEV with <strong>14 questions attached at the same time</strong>: which department, is it urgent, what’s the sentiment, what language, and more. All fourteen came back together, and the cost didn’t move.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-A3IZaCYg3P7mqN-isnySA.png" /><figcaption>Questions?</figcaption></figure><p>That only works because the questions are processed in parallel against the same prefilled context, rather than as separate sequential calls. It’s a small architectural choice with a large practical consequence: you stop rationing your questions. If you can think of a useful thing to ask, you just ask it.</p><h3>Turning “how sure are you” into actual code</h3><p>The feature I keep coming back to, because I think it’s the most underrated one, is the confidence score — and specifically, that RLCD training makes it <em>honest</em>. If JEV tells you it’s 90% confident, that’s meant to hold up statistically: 90% of the time it says that, it should be right. LLMs are notorious for faking confidence; this is a model explicitly trained not to.</p><p>That honesty turns a probability into a business rule.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Y3o16iqcm06oa4n79j0urw.png" /><figcaption>confidence score</figcaption></figure><p>Above 90%, you act without a human in the loop. Between 60 and 90, your code can ask a clarifying follow-up. Below 60, you route it to a person. That’s not a prompt — that’s an if statement, and it&#39;s the exact sentence TypeSafe uses to describe their own ambition: <strong>AI moving from a feature to a primitive.</strong></p><h3>I rebuilt a Flipkart feature in under 30 lines</h3><p>To actually feel this instead of just reading about it, I picked something I use constantly: those per-attribute ratings you see on a Flipkart product page — camera rated 4.6, battery 4.1, display 4.8 — each pulled straight out of thousands of raw text reviews. That’s normally a fine-tuned NLP model, built by a data science team, on their own labelled data.</p><p>I fed 50 synthetic reviews for a hypothetical phone into JEV, asked it two questions per attribute across seven attributes — <em>“does this review mention the camera?”</em> and <em>“how satisfied does the reviewer sound about it?”</em> — for every single review, and let it aggregate.</p><p>No training. No fine-tuning. No labelled dataset of my own. Just an API call and 50 reviews, and out came the same kind of attribute-level rating breakdown Flipkart’s engineering team almost certainly built with a custom model. It took me under half an hour end to end, including reading the docs.</p><p>That, more than any benchmark number, is what convinced me this isn’t just hype dressed up as a product launch.</p><h3>Where I think this actually goes</h3><p>I don’t think JEV replaces LLMs, and neither does TypeSafe — they’re explicit that this is a <em>system-1</em> model, meaning it isn’t especially “intelligent” by design. You wouldn’t ask it to write your essay or plan your product roadmap. What I do think is coming:</p><ul><li><strong>Agentic routing gets cheap.</strong> Every “which tool should I call” or “is this step safe” decision inside an AI agent is a natural fit — the heavy reasoning stays with the LLM, the thousand small decisions around it move to something like JEV.</li><li><strong>AI quietly becomes plumbing, not a headline feature.</strong> The way databases or caches are invisible infrastructure today, fast decision models start disappearing into the background of ordinary software.</li><li><strong>A new job description shows up.</strong> Someone whose entire role is deciding <em>which</em> decisions get handed to a model like this, and what confidence thresholds should gate them.</li><li><strong>The field gets crowded fast.</strong> A benchmark called JEV-Bench appeared within a week of launch, and it’s already tracking a dozen copycat and open-source alternatives — because the recipe (take a pretrained model, swap the decode head, train on synthetic data) isn’t expensive to replicate.</li></ul><h3>And where it still falls short</h3><p>I’d be doing you a disservice if I only told you the good half.</p><p>There’s <strong>no independent, trustworthy benchmark yet</strong> — every number so far comes from TypeSafe itself. It’s <strong>text-only</strong> for now, so anything visual or audio-based is out of reach. It has <strong>no web access</strong>, so it’s blind to anything past its training cutoff. It gives you <strong>zero explanation</strong> for why it picked an answer — just a score, no reasoning trail — which rules it out anywhere you’d need to justify a decision after the fact, like lending or compliance. And it’s explicitly <strong>not a highly intelligent model</strong> — it’s tuned for fast, narrow calls, not nuanced judgment. It’s also closed source, with no paper and no dataset, which is its own kind of risk if a well-funded lab or a strong open alternative catches up.</p><h3>Where I’ve landed</h3><p>The honest version of my takeaway is this: <strong>JEV isn’t a smarter model. It’s a smaller, faster, cheaper <em>decision</em>, wearing a transformer’s clothing.</strong> And once decisions get cheap enough to stop rationing, engineers start finding decisions to make everywhere — in the UI, in the pipeline, in places we never thought to ask AI a question before, because asking used to cost too much, in both time and money.</p><p>That’s the actual shift I sat down to write about. Not another chatbot. A quiet, structural change in what “using AI” is even going to mean, a year or two from now.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5f6645c01df5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/android-knowledge-store/what-is-jev-inside-the-ai-model-built-to-replace-llms-for-fast-decisions-5f6645c01df5">What Is JEV? Inside the AI Model Built to Replace LLMs for Fast Decisions</a> was originally published in <a href="https://medium.com/android-knowledge-store">Tech Nexus</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[SocketCAN and Virtual CAN: The Bus That Doesn’t Exist]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/android-knowledge-store/socketcan-and-virtual-can-the-bus-that-doesnt-exist-933174810ce2?source=rss-779303a3099------2"><img src="https://cdn-images-1.medium.com/max/1920/1*d-oQD4LuTEiFvd1jmuViEg.png" width="1920"></a></p><p class="medium-feed-snippet">I had already spent close to two years working on IVI and cluster programs , one day, when a junior on my team asked me a very simple&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/android-knowledge-store/socketcan-and-virtual-can-the-bus-that-doesnt-exist-933174810ce2?source=rss-779303a3099------2">Continue reading on Tech Nexus »</a></p></div>]]></description>
            <link>https://medium.com/android-knowledge-store/socketcan-and-virtual-can-the-bus-that-doesnt-exist-933174810ce2?source=rss-779303a3099------2</link>
            <guid isPermaLink="false">https://medium.com/p/933174810ce2</guid>
            <category><![CDATA[linux]]></category>
            <category><![CDATA[can]]></category>
            <category><![CDATA[socket-programming]]></category>
            <category><![CDATA[automotive]]></category>
            <category><![CDATA[socketcan]]></category>
            <dc:creator><![CDATA[Budhdi Sharma]]></dc:creator>
            <pubDate>Mon, 14 Sep 2026 07:11:33 GMT</pubDate>
            <atom:updated>2026-09-14T07:11:33.014Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Industry Wake-Up: Secure Multi Agent Architecture]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/autosec-journal/the-industry-wake-up-secure-multi-agent-architecture-4f9f99d4bd03?source=rss-779303a3099------2"><img src="https://cdn-images-1.medium.com/max/1024/1*67JwA1W3OeWOyT4RKag9Pg.png" width="1024"></a></p><p class="medium-feed-snippet">There&#x2019;s a particular kind of meeting that happens in engineering teams right now. Someone demos an AI agent, it browses the web, reads a&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/autosec-journal/the-industry-wake-up-secure-multi-agent-architecture-4f9f99d4bd03?source=rss-779303a3099------2">Continue reading on AIDefense Log »</a></p></div>]]></description>
            <link>https://medium.com/autosec-journal/the-industry-wake-up-secure-multi-agent-architecture-4f9f99d4bd03?source=rss-779303a3099------2</link>
            <guid isPermaLink="false">https://medium.com/p/4f9f99d4bd03</guid>
            <category><![CDATA[agentic-ai-architecture]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[ai-cyber-security]]></category>
            <category><![CDATA[ai-security]]></category>
            <dc:creator><![CDATA[Budhdi Sharma]]></dc:creator>
            <pubDate>Tue, 09 Jun 2026 00:38:51 GMT</pubDate>
            <atom:updated>2026-06-09T00:38:51.056Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Brilliant Assistant or Security Nightmare? The 6 Hidden Risks of OpenClaw]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/autosec-journal/brilliant-assistant-or-security-nightmare-the-6-hidden-risks-of-openclaw-732c483ffa56?source=rss-779303a3099------2"><img src="https://cdn-images-1.medium.com/max/1024/1*D3NndNYYOTzsD_tAcFdSVw.png" width="1024"></a></p><p class="medium-feed-snippet">Imagine you hired a new assistant. Brilliant. Fast. Never sleeps. You hand them your house keys, your office keycard, your email password&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/autosec-journal/brilliant-assistant-or-security-nightmare-the-6-hidden-risks-of-openclaw-732c483ffa56?source=rss-779303a3099------2">Continue reading on AIDefense Log »</a></p></div>]]></description>
            <link>https://medium.com/autosec-journal/brilliant-assistant-or-security-nightmare-the-6-hidden-risks-of-openclaw-732c483ffa56?source=rss-779303a3099------2</link>
            <guid isPermaLink="false">https://medium.com/p/732c483ffa56</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[ai-agent-security]]></category>
            <category><![CDATA[ai-cyber-security]]></category>
            <category><![CDATA[ai-security]]></category>
            <category><![CDATA[openclaw]]></category>
            <dc:creator><![CDATA[Budhdi Sharma]]></dc:creator>
            <pubDate>Mon, 08 Jun 2026 23:48:05 GMT</pubDate>
            <atom:updated>2026-06-08T23:48:05.083Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[What If AOSP/AAOS Had a Brain?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/android-knowledge-store/what-if-aosp-aaos-had-a-brain-af46761ea10f?source=rss-779303a3099------2"><img src="https://cdn-images-1.medium.com/max/1600/1*Z9j3SxxXzUKy9G3QQQ2BdA.png" width="1600"></a></p><p class="medium-feed-snippet">A tiny ML model, running on the NPU, watching everything &#x2014; and fixing it before you notice.</p><p class="medium-feed-link"><a href="https://medium.com/android-knowledge-store/what-if-aosp-aaos-had-a-brain-af46761ea10f?source=rss-779303a3099------2">Continue reading on Tech Nexus »</a></p></div>]]></description>
            <link>https://medium.com/android-knowledge-store/what-if-aosp-aaos-had-a-brain-af46761ea10f?source=rss-779303a3099------2</link>
            <guid isPermaLink="false">https://medium.com/p/af46761ea10f</guid>
            <category><![CDATA[aosp]]></category>
            <category><![CDATA[androiddev]]></category>
            <category><![CDATA[android]]></category>
            <category><![CDATA[ai-model]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Budhdi Sharma]]></dc:creator>
            <pubDate>Fri, 06 Mar 2026 12:12:35 GMT</pubDate>
            <atom:updated>2026-03-06T12:12:35.585Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Day 3: Adversarial Attacks in Depth: FGSM, PGD, C&W, and Why the Obvious Defenses Fail]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/autosec-journal/day-3-adversarial-attacks-in-depth-fgsm-pgd-c-w-and-why-the-obvious-defenses-fail-89b01c8fa123?source=rss-779303a3099------2"><img src="https://cdn-images-1.medium.com/max/1344/1*YFXRYZN8nF7ZRwxM2lxyZQ.jpeg" width="1344"></a></p><p class="medium-feed-snippet">Day 2 gave you the math. Today we use it. Real attacks. Real numbers. Real code. And the uncomfortable truth about why defending against&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/autosec-journal/day-3-adversarial-attacks-in-depth-fgsm-pgd-c-w-and-why-the-obvious-defenses-fail-89b01c8fa123?source=rss-779303a3099------2">Continue reading on AIDefense Log »</a></p></div>]]></description>
            <link>https://medium.com/autosec-journal/day-3-adversarial-attacks-in-depth-fgsm-pgd-c-w-and-why-the-obvious-defenses-fail-89b01c8fa123?source=rss-779303a3099------2</link>
            <guid isPermaLink="false">https://medium.com/p/89b01c8fa123</guid>
            <category><![CDATA[security]]></category>
            <category><![CDATA[ai-security]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[ai-cyber-security]]></category>
            <dc:creator><![CDATA[Budhdi Sharma]]></dc:creator>
            <pubDate>Mon, 02 Mar 2026 05:17:50 GMT</pubDate>
            <atom:updated>2026-03-02T05:17:50.754Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Threat Model That Actually Protects a Connected Car]]></title>
            <link>https://medium.com/securedrive-systems/the-threat-model-that-actually-protects-a-connected-car-e636c134bb65?source=rss-779303a3099------2</link>
            <guid isPermaLink="false">https://medium.com/p/e636c134bb65</guid>
            <category><![CDATA[iso-21434]]></category>
            <category><![CDATA[automotive-cyber-security]]></category>
            <category><![CDATA[tara]]></category>
            <category><![CDATA[automotive-cybersecurity]]></category>
            <category><![CDATA[automotive]]></category>
            <dc:creator><![CDATA[Budhdi Sharma]]></dc:creator>
            <pubDate>Fri, 27 Feb 2026 17:46:44 GMT</pubDate>
            <atom:updated>2026-02-27T17:46:44.053Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zYF4DMu3nyI1kZvAM9Fq_w.jpeg" /></figure><p>Everyone says they do TARA. Very few do it in a way that changes anything. Here’s what it looks like when it’s done right, and the real cost when it isn’t. This is Part 04 in a series about the TARA in Practice</p><blockquote>Early in my career, I sat in a TARA review where a team confidently presented 47 threat scenarios, each one perfectly structured, each with a mitigation assigned. It was a beautiful document. The problem? The architecture it described had been frozen six months earlier. Nobody changed anything. The TARA was written to describe decisions already made, not to drive new ones. That document didn’t protect a single vehicle. It just gave the program a stamp.</blockquote><p>TARA — Threat Analysis and Risk Assessment — is the analytical backbone of ISO/SAE 21434. Done right, it’s the moment where cybersecurity stops being abstract and becomes a set of specific engineering decisions with specific justifications. Done wrong, it’s the most expensive document in your program that changes absolutely nothing.</p><p>This final article in the series gets into the mechanics. Not the theory — the actual practice. What a real TARA looks like, where it fails, what it’s supposed to produce, and what that means for the engineers building the system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kNdofv_BRegLWbisRy5wwg.png" /><figcaption>TARA Loop</figcaption></figure><h3>Step by Step: How a Real TARA Actually Works</h3><p>Let me walk through TARA using a concrete vehicle system — a telematics control unit (TCU). It’s a real ECU, present in virtually every connected vehicle today, and it’s one of the highest-risk components in the vehicle because it faces the internet directly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d6TWBsbtDDNpUHpQLHuMzw.png" /><figcaption>Item Defination</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qdCDZbsqxvWew56unofMnQ.png" /><figcaption>Asset</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cpT7RBilrAvQzWjJ8JtSMQ.png" /><figcaption>Threats</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mLjPjUzoeVgFm5QYu7q8qw.png" /><figcaption>Risks</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e4yPYuinoame_sCkZbjhhg.png" /><figcaption>Treatment</figcaption></figure><h3>A Real TARA Output — The TCU Threat Table</h3><p>Here’s what a slice of a properly executed TARA actually looks like. Not a comprehensive document — just five rows covering the TCU’s highest-risk threat scenarios. Every field earned, none copied and pasted from a template.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PcoXyq-mO0uoWoyqni8uwg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O9Lp4hLZMhX0FErW4JZ6RQ.png" /></figure><h3>The Risk Matrix, Reading It Properly</h3><p>The risk rating isn’t just a number you pick. It’s the product of two independent assessments — impact and attack feasibility — that together determine how much engineering rigor your mitigation needs. Here’s how the matrix works:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rGSbIs2cxis1ZkiTNUgvkQ.png" /><figcaption>CAL 4 conversion</figcaption></figure><h3>The Supply Chain Problem Nobody Wants to Own</h3><p>Here’s the uncomfortable truth about TARA in a real vehicle program: you’re not just doing it for your own components. ISO/SAE 21434 requires cybersecurity obligations to flow down through every tier of your supply chain. And most programs handle this badly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1NF8gG2-wienSlhxIG05Ug.png" /><figcaption>Supply Chain Threat Protection</figcaption></figure><p>The Tier-2 gap is real and persistent. An OEM can have a beautifully certified CSMS with a comprehensive TARA at the system level — and be running a modem chip firmware with unpatched CVEs because nobody contractually required the silicon vendor to maintain a vulnerability disclosure program. That’s a supply chain attack vector that no amount of system-level security architecture will close.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aVjTw0WKQTR9GkQjB_FRyA.png" /></figure><blockquote>The TARA for the component you built perfectly doesn’t protect you from the library you didn’t audit.</blockquote><blockquote>— Hard-learned lesson, multiple program reviews</blockquote><h3>The Three Patterns That Kill TARA Effectiveness</h3><p>After years of seeing these documents produced across different programs and different organizations, the failure modes cluster into three distinct patterns. Each one produces a TARA that passes the audit and protects nothing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AaGOWmb-nD4YnuYmaO9Xng.png" /><figcaption>The test for tara</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-8OXO9PKRRLXptVYd3qkSw.png" /><figcaption>TARA Findings</figcaption></figure><h3>Where This Series Leaves You</h3><p>We started this series with a journalist losing control of a Jeep at 70mph because two researchers were sitting in a living room. We’ve covered how ECU boot chains work and where they fail, what HSMs actually do versus what teams think they do, how ISO/SAE 21434 reshapes the entire engineering process, and now — how the threat modeling at the center of that standard needs to work to produce real security rather than real documentation.</p><p>The common thread across all four articles is this: connected vehicle security isn’t hard because the technical solutions don’t exist. The cryptographic primitives work. The HSMs are available. The standards are written. It’s hard because it requires organizational discipline, cross-functional commitment, and the willingness to let security findings change expensive engineering decisions.</p><p>That last part — the willingness to change decisions — is where programs succeed or fail. A TARA that changes nothing is a liability, not an asset. Secure boot that gets bypassed by a debug flag is theatrical. ISO 21434 compliance that lives in documents but not in architectures is a regulatory checkbox with no security value.</p><p>The vehicles being designed today will be on roads in 2027 and running until 2042. The threat landscape in 2042 is unknowable. What’s knowable is whether the engineering discipline exists to adapt to it — and that discipline starts with doing the foundational work properly now. Not for the auditor. For the driver.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e636c134bb65" width="1" height="1" alt=""><hr><p><a href="https://medium.com/securedrive-systems/the-threat-model-that-actually-protects-a-connected-car-e636c134bb65">The Threat Model That Actually Protects a Connected Car</a> was originally published in <a href="https://medium.com/securedrive-systems">SecureDrive Systems</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ISO/SAE 21434: The Standard That’s Quietly Reshaping Every Car on the Road]]></title>
            <link>https://medium.com/securedrive-systems/iso-sae-21434-the-standard-thats-quietly-reshaping-every-car-on-the-road-56bed0d92c27?source=rss-779303a3099------2</link>
            <guid isPermaLink="false">https://medium.com/p/56bed0d92c27</guid>
            <category><![CDATA[automotive]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[tara]]></category>
            <category><![CDATA[iso-21434]]></category>
            <category><![CDATA[automotive-cybersecurity]]></category>
            <dc:creator><![CDATA[Budhdi Sharma]]></dc:creator>
            <pubDate>Fri, 27 Feb 2026 17:01:24 GMT</pubDate>
            <atom:updated>2026-02-27T17:01:24.329Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>It’s not just another compliance document. It’s a complete rethink of how the automotive industry is supposed to build things that can’t be broken into.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_ihbPqe_bnMydObEuPtEOQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cDbpoTTkxgvZumZUtrzK7Q.png" /></figure><p>Let me tell you something that still surprises people when I bring it up in engineering reviews. The average new vehicle today contains somewhere between 100 and 150 million lines of code. A modern fighter jet runs on roughly 25 million. We’ve built software systems into our cars that dwarf military aircraft — and for most of automotive history, we treated security as an afterthought.</p><p>That era is ending. And the document accelerating that end is called ISO/SAE 21434.</p><p>I’ve been living with this standard across OEM and Tier-1 programs for years now — applying it, arguing about it in review rooms, watching suppliers scramble to meet it and occasionally watching them pretend to meet it. What I want to give you here isn’t a summary of the clauses. There are plenty of those online. What I want to give you is the actual story of what this standard does, why it matters, and what it looks like when it’s done right versus when it’s done for appearances.</p><h3>First — Why Does a Car Even Need a Cybersecurity Standard?</h3><p>To understand why ISO/SAE 21434 exists, you need to understand what happened to cars over the past fifteen years. They got connected. Then, more connected. Then they started updating themselves over the air, talking to infrastructure, running machine learning models, and integrating deeply with cloud backends.</p><p>Each one of those capabilities is genuinely useful. Each one is also a door that didn’t exist before.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o0mfWNtkRW83I-D_ya2EXA.png" /></figure><blockquote>In many years of working across OEM and Tier-1 programs, I’ve seen the same realization hit teams at different times. They build something sophisticated — a connected telematics unit, an OTA manager, a gateway ECU — and then someone asks: “What happens if an attacker gets in here?” And there’s a pause. A long one. Because the answer wasn’t designed in from the beginning. It was going to be bolted on at the end. ISO/SAE 21434 exists specifically to make that pause happen at architecture time, not at the end of development when changing anything costs a fortune.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KSvLWQVjvRuRMbDGjEsCpw.png" /></figure><h3>What ISO/SAE 21434 Actually Does</h3><p>The standard’s full name — <em>Road Vehicles — Cybersecurity Engineering</em> — gives you the core idea. It’s an engineering standard. Not a compliance checklist, not a documentation exercise. It defines how organizations must engineer cybersecurity into vehicles across their entire existence, from the first sketch on a whiteboard through the day the car gets scrapped.</p><p>Three things make it different from what the industry did before:</p><p><strong>It’s risk-based, not rule-based.</strong> Instead of saying “you must use encryption algorithm X,” it says you must identify your threats, assess the risk each poses, and implement mitigations proportional to that risk. The mitigations are your engineering call. The rigor of the process is the requirement.</p><p><strong>It covers the entire lifecycle.</strong> Most engineering standards focus on development. 21434 extends all the way through production, in-field operation, incident response, and end of life. A vehicle sold today may be on the road for fifteen years. The cybersecurity obligations don’t end at the factory gate.</p><p><strong>It flows through the supply chain.</strong> OEMs can’t just do this work themselves and call it done. The standard requires cybersecurity obligations to be flowed down to every tier of supplier. A Tier-2 silicon vendor has to demonstrate they’ve done security engineering on their component. That’s a significant shift.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iGuQaFKWj-1KQ0j-mUHKbw.png" /></figure><h3>The Engine at the Heart of It: TARA</h3><p>If I had to pick the single most important concept in 21434, it’s TARA — Threat Analysis and Risk Assessment. Every security decision in a compliant program traces back to a TARA. And the quality of your TARA determines the quality of everything that follows.</p><p>A TARA isn’t complicated in concept. You identify what’s valuable in your system (the assets). You figure out what bad things could happen to those assets (the threats). You estimate how likely and how damaging each threat is (the risk). And then you design mitigations proportional to the risk level you’ve assessed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8S9DnNwKrkRjmMugubCY7w.png" /></figure><p>Where teams get TARA wrong is treating it as a documentation exercise rather than an engineering activity. A TARA that was written to pass an audit — where someone reverse-engineered mitigations to match an architecture that was already frozen — is useless. The value is in doing it before the architecture is frozen, when the findings can actually change design decisions.</p><h3>The Regulatory Picture — Why This Is Now Mandatory</h3><p>ISO/SAE 21434 is a voluntary standard — but UNECE Regulation 155 is not. R155 mandates that automotive OEMs operating in regulated markets maintain a certified Cybersecurity Management System (CSMS). And the reference framework for what a CSMS looks like? ISO/SAE 21434. So in practice, if you want to sell vehicles in Europe, Japan, South Korea, or the 54 markets that have adopted R155, you need to implement 21434.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wmZeK4rqQHRTX--bqpHBsQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*s56WxGSurgHNjdY5mXaPog.png" /></figure><h3>What Good Looks Like vs. What Teams Actually Do</h3><p>This is the part I never see written about, so let me be direct about it. There’s a version of 21434 compliance that’s genuinely useful and a version that’s theater. I’ve seen both.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_-z6L5RNhEWa73nBRBfONA.png" /></figure><h3>The Autonomy Dimension — Where It Gets More Serious</h3><p>Everything above applies to any connected vehicle. But when you add autonomy to the picture, the stakes change in a specific way. A cyberattack on a conventional vehicle’s infotainment system is a bad day. A cyberattack on an autonomous vehicle’s perception stack could be a fatal collision.</p><p>This is not hypothetical engineering concern territory — researchers have already demonstrated GNSS spoofing attacks that mislead vehicle navigation, camera-based attacks using adversarial patterns printed on road signs, and LiDAR spoofing that inserts phantom objects into a vehicle’s perception space. These attacks don’t exploit software vulnerabilities in the traditional sense. They corrupt what the AI sees.</p><p>ISO/SAE 21434 scopes into this territory by requiring that autonomy and perception pipelines be included in TARA — their data inputs, their model integrity, and their update channels all have to be threat-modeled. This is work the industry is genuinely just beginning to grapple with. It’s also an area I’m actively researching: adversarial ML threat detection in embedded edge environments is one of the most technically interesting open problems sitting at the intersection of automotive and AI security right now.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gtp3XMaxblEKxzwgN4PGdQ.png" /></figure><h3>What This Means if You’re Building Automotive Software Today</h3><p>Whether you’re at an OEM, a Tier-1, or a software supplier, the practical implications of 21434 are the same: your customers will ask you for evidence of cybersecurity engineering, not just a claim that you take security seriously. Evidence means documented TARAs, cybersecurity requirements traced to design decisions, verification test results, and a CSMS you can actually explain when someone audits it.</p><p>The teams that are ahead of this are the ones that started treating cybersecurity as an engineering discipline — with the same rigor as functional safety — rather than as a compliance requirement to be managed. The teams that are behind are the ones still trying to retrofit security onto architectures that were designed without it.</p><p>The standard won’t make building secure vehicles easy. The attack surface is genuinely complex, the threat landscape is evolving faster than most development cycles, and the supply chain dependencies run deep. But it gives the industry a shared language and a shared process for approaching the problem systematically. That matters enormously when you’re coordinating work across hundreds of suppliers and multiple vehicle generations.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=56bed0d92c27" width="1" height="1" alt=""><hr><p><a href="https://medium.com/securedrive-systems/iso-sae-21434-the-standard-thats-quietly-reshaping-every-car-on-the-road-56bed0d92c27">ISO/SAE 21434: The Standard That’s Quietly Reshaping Every Car on the Road</a> was originally published in <a href="https://medium.com/securedrive-systems">SecureDrive Systems</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Inside the ECU: Secure Boot, HSMs,
and Why Most Implementations
Get It Wrong]]></title>
            <link>https://medium.com/securedrive-systems/inside-the-ecu-secure-boot-hsms-and-why-most-implementations-get-it-wrong-378933ed3d20?source=rss-779303a3099------2</link>
            <guid isPermaLink="false">https://medium.com/p/378933ed3d20</guid>
            <category><![CDATA[secure-boot]]></category>
            <category><![CDATA[unece-r155]]></category>
            <category><![CDATA[hms]]></category>
            <category><![CDATA[automotive-cybersecurity]]></category>
            <category><![CDATA[iso-21434]]></category>
            <dc:creator><![CDATA[Budhdi Sharma]]></dc:creator>
            <pubDate>Thu, 26 Feb 2026 16:12:35 GMT</pubDate>
            <atom:updated>2026-02-26T16:12:35.069Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>When people outside the automotive industry hear “ECU cybersecurity,” they tend to imagine it as a software problem — write better code, patch the bugs, done. Having worked across multiple vehicle programs, I can tell you it’s a fundamentally different kind of challenge. The adversarial surface lives at the intersection of hardware, firmware, networking, and manufacturing — and the security decisions you make at architecture time are nearly impossible to undo once a vehicle is in production.</blockquote><p>Let’s get into it properly.</p><h3>The Boot Chain — Where Trust Either Starts or Dies</h3><p>Every ECU in a modern vehicle runs through a boot sequence before it does anything useful. The security of that sequence is the single most foundational decision in automotive cybersecurity. Get it right and you have a hardware-anchored trust foundation that every subsequent security measure can build on. Get it wrong and everything else — your network segmentation, your encryption, your intrusion detection — is a house of cards.</p><p>Here’s what a typical automotive ECU boot chain looks like, and what each stage is actually responsible for:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MraPSb9rpEKczWHX2CcvvA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*f-hUo6pN1rQ9iQ2YSyuLFQ.png" /></figure><p>The critical concept here is that each stage verifies the next, it doesn’t just assume the next stage is fine. This “chain of trust” means a compromise has to break through multiple cryptographic verification steps to reach execution. That’s not unbreakable, but it raises the cost of an attack dramatically.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WTM5E4uTHiaBMoUn6p8LSA.png" /></figure><h3>Hardware Security Modules — What They Actually Do</h3><p>The HSM might be the most misunderstood component in automotive security discussions. I’ve sat in architecture reviews where the team confidently states “we have an HSM” as if that’s a complete security answer. It’s not. An HSM is a tool. How you use it — specifically what you put inside it and what you keep outside it — determines everything.</p><p>The fundamental value of an HSM is isolation. It provides a physically separate execution environment where cryptographic operations happen. Even if an attacker has fully compromised your main application processor — root access, arbitrary code execution, everything — they cannot reach into the HSM and extract the keys stored there. The HSM is a vault that communicates with the rest of the system through a narrow, controlled interface.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4dKv-jjQCpMBSV2aYAyDUg.png" /><figcaption>HSM_Architecture_Automotive SoC Security Boundary</figcaption></figure><p>In practice, most automotive HSMs implement either the SHE (Secure Hardware Extension) standard — common in cost-sensitive microcontrollers — or the more capable EVITA profile for gateway and central compute ECUs. SHE is simpler and cheaper; EVITA gives you asymmetric cryptography, which matters enormously for OTA update authentication and inter-ECU certificate-based trust.</p><p>Renesas R-Car SoCs and RH850 MCUs — both widely deployed across OEM programs I’ve worked with — ship with embedded HSMs optimized for gateway protection, secure external communication, and in-vehicle network security. Vector’s MICROSAR HSM firmware layered on top provides an ISO/SAE 21434-certified implementation. That certification matters in practice: it saves TARA effort on the HSM component itself, letting teams focus on analysis where it actually needs to go.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4G-ayvZotBAf5dCDlmaWWA.png" /></figure><h3>OTA Security — The Update Mechanism That Must Never Fail</h3><p>Over-the-air firmware updates have transformed how the industry thinks about vehicle lifecycles. They’ve also introduced one of the highest-severity attack vectors in the entire automotive threat landscape. If someone can push arbitrary firmware through your OTA channel, they don’t need to exploit any other vulnerability — they own every vehicle in your fleet simultaneously.</p><blockquote>A compromised OTA pipeline isn’t a single vehicle incident. It’s a fleet-wide incident measured in millions of vehicles. The blast radius is unlike anything else in automotive security.</blockquote><blockquote>— From production architecture review experience</blockquote><p>The SAFUA framework and Adaptive AUTOSAR’s update manager give us the conceptual building blocks. But understanding how these pieces connect in production is what matters. Here’s the secure OTA verification pipeline that a properly implemented system follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DI4PB7mY0Bis41i6ZDBg3A.png" /></figure><p>The monotonic counter step is one that teams consistently underinvest in. Its job is simple: prevent an attacker from rolling back firmware to an older, known-vulnerable version by replaying a legitimate signed package from the past. Without a monotonic counter stored inside the HSM that software can’t reset it, and your entire OTA signature verification is vulnerable. I’ve had the conversation too many times — “the attacker would need the old package” — yes, and old packages are not difficult to obtain if you know where to look.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cnKpP75At53r0wNJg3wl4w.png" /></figure><h3>Standards Compliance — What ISO 21434 Actually Demands Here</h3><p>I want to be direct about something: a lot of the ISO/SAE 21434 discussion in the industry treats compliance as the goal. It shouldn’t be. The standard exists to describe what good security engineering looks like. If you’re doing the engineering properly, compliance follows. If you’re checking boxes to get the audit done, you’ll pass the audit and ship an insecure vehicle.</p><p>That said, understanding precisely what the standard requires for secure boot and HSM implementation is genuinely useful — both for engineering teams and for the TARA work that accompanies it. Here’s how the key elements map against typical program reality:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UnnI35GO-mX6dip167SLTg.png" /></figure><p>The debug interface lockdown row is the one that makes people uncomfortable in review meetings. Production ECUs with accessible JTAG ports are more common than they should be — not because engineers don’t know better, but because reliably disabling them across an entire production run is operationally complex and sometimes deprioritized in the final push to SOP. Physical JTAG access to a production ECU bypasses most software-level protections entirely. It needs to be in the TARA explicitly, and the mitigation needs to be verified at production validation, not just assumed.</p><h3>What Actually Gets Missed — The Real Implementation Checklist</h3><p>After years of architecture reviews, supplier audits, and production integrations, here are the checks that tell me whether a team has been thinking about this seriously or just fulfilling a process requirement. The partial items are where real programs consistently fall short.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*722VJo4DZdXH82uoZsFFpA.png" /></figure><p>The partial items have a pattern — they’re not technically difficult, but they require coordination across embedded engineering, manufacturing, and security that doesn’t always happen naturally. That organizational friction is frequently the real security gap in automotive programs. The engineering knowledge exists. The process discipline to apply it consistently is where things break down.</p><h3>Where This Is Heading — Isolated AI Compute and What It Changes</h3><p>High-performance compute platforms in vehicles — NVIDIA DRIVE, Qualcomm Snapdragon Ride, Arm’s AE-profile SoCs — bring a new dimension to this conversation. These platforms run perception models, sensor fusion, and increasingly AI-based decision-making. Having worked extensively on Qualcomm XR SoCs with SLAM tracking and sensor pipelines, the architecture parallels are real: the same principles of hardware isolation, trusted execution environments, and authenticated inter-domain communication that matter in AR/VR platforms matter here too. The stakes are simply different.</p><p>A compromised SLAM model crashing AR glasses is inconvenient. A compromised perception model in an ADAS system is a safety incident with legal and human consequences.</p><p>Arm’s Zena CSS specifically addresses this by providing isolated AI compute environments — hardware-separated domains where AI inference workloads run independently from safety-critical control software. The layered defense-in-depth model this enables means that a compromised AI inference path cannot directly influence the safety-critical control path without crossing a hardware isolation boundary with its own authentication requirements.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YSw9gvmams4RfK4EM0GNPA.png" /></figure><p>The short version for architects working on this now: design your AI compute domains with the same isolation discipline you’d apply to any safety-critical partition. Define the trust boundary clearly. Authenticate inputs to the AI stack where possible. And include AI inference pipelines in your TARA scope explicitly — because if they’re not in scope, no one is reasoning systematically about what happens when they’re attacked.</p><p>The fundamentals here — chain of trust, HSM isolation discipline, OTA anti-rollback, debug hygiene — are not glamorous topics. They don’t make for exciting conference talks. But in my experience, they’re exactly what separates vehicles that are genuinely secure from vehicles that merely appear secure until someone looks carefully. And in this industry, someone always looks carefully eventually. Better that someone is on your own team, during development, than a researcher with a public disclosure deadline and a press contact.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=378933ed3d20" width="1" height="1" alt=""><hr><p><a href="https://medium.com/securedrive-systems/inside-the-ecu-secure-boot-hsms-and-why-most-implementations-get-it-wrong-378933ed3d20">Inside the ECU: Secure Boot, HSMs,
and Why Most Implementations
Get It Wrong</a> was originally published in <a href="https://medium.com/securedrive-systems">SecureDrive Systems</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When Your Car Gets Hacked: The Real World of Automotive Cybersecurity]]></title>
            <link>https://medium.com/securedrive-systems/when-your-car-gets-hacked-the-real-world-of-automotive-cybersecurity-3da19a30b76e?source=rss-779303a3099------2</link>
            <guid isPermaLink="false">https://medium.com/p/3da19a30b76e</guid>
            <category><![CDATA[automotive-cybersecurity]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[automotive]]></category>
            <category><![CDATA[hacking]]></category>
            <dc:creator><![CDATA[Budhdi Sharma]]></dc:creator>
            <pubDate>Wed, 25 Feb 2026 17:31:12 GMT</pubDate>
            <atom:updated>2026-02-25T17:41:35.904Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w5Zp_NTqo5OQC_AJoop23A.jpeg" /><figcaption>CAR CS</figcaption></figure><p>Modern vehicles are rolling computers. And like every computer ever made, they can be broken into. Here’s what that really looks like, and why the industry is scrambling to catch up.</p><blockquote>Picture this: you’re driving on the highway at 70 mph when your radio starts blaring, your windshield wipers flick on, and then — the steering wheel jerks out of your control. You didn’t do anything. The car just… did it. On its own. Because someone miles away, sitting at a laptop, decided to try something.</blockquote><p>This is not a thriller novel. In 2015, it actually happened to a journalist who volunteered to drive a Jeep Cherokee while two security researchers took control of the vehicle remotely over the internet. The world watched, and the automotive industry has never quite been the same since.</p><p>We’re now deep into an era where the average new car has more lines of software code than a commercial aircraft. These vehicles connect to the internet, talk to each other, receive over-the-air updates, and integrate with your smartphone. That connectivity is genuinely wonderful — but it also opens a door that didn’t exist when cars were purely mechanical.</p><p>This is the first in a series where we’ll dig into automotive cybersecurity properly. Not with surface-level summaries, but with the real stuff: how attacks work, what the regulations actually require, and what engineers in this space deal with every day. Let’s start at the beginning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UcWoiXISfnGheYk0dqUzFw.png" /><figcaption>Vehicle Threat</figcaption></figure><h3>The Hack That Changed Everything</h3><p>You have to understand the 2015 Jeep Cherokee incident in its full context to appreciate why it mattered so much. Before that summer, most people inside and outside the auto industry treated vehicle cybersecurity as a theoretical problem. Interesting to think about, perhaps — but not exactly urgent.</p><p>Researchers Charlie Miller (who’d previously hacked iPhones) and Chris Valasek (a longtime car security enthusiast) had been poking at vehicle systems for a couple of years. They’d done earlier demos requiring physical access to the car. But the 2015 attack was something else entirely. It required no physical access whatsoever. Miller was sitting in his living room in St. Louis while the Jeep’s victim-driver, <em>Wired</em> journalist Andy Greenberg, drove on a highway outside the city.</p><p>The entry point was the Jeep’s Uconnect entertainment system, which connected to the Sprint cellular network. It had a publicly routable IP address and a software vulnerability that allowed the researchers to push new firmware to a chip in the head unit. From there, they could communicate with the vehicle’s internal CAN bus — the nervous system that connects nearly every electronic component in a modern car.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QJ1WHqFCxELIiqjQ6xxppg.png" /><figcaption>Jeep Exploit</figcaption></figure><p>What made this attack genuinely alarming wasn’t just that it worked — it was the breadth of what the researchers could do once they were in. They could kill the engine, disable the brakes at low speeds, turn the steering wheel (at reverse speeds), crank the AC, blast the radio, and activate the windshield wipers. All remotely. All while the driver sat helplessly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Gen1NlRtOJ9iqomrU9O_oA.png" /><figcaption>Valasek</figcaption></figure><p>The CAN bus problem at the heart of this exploit is worth understanding. The CAN protocol was designed decades ago specifically to allow the many electronic control units in a vehicle to talk to each other efficiently. It’s elegant engineering, but it was built in an era when no one imagined a car’s radio would have an internet connection that could reach the brakes. There’s essentially no authentication on the CAN bus. If you can get a message onto it, the receiving component will generally trust it and obey.</p><p>This is the foundational tension in automotive cybersecurity: legacy systems built for reliability in isolated environments are now connected to a world full of adversarial actors.</p><h3>Tesla’s Approach: Move Fast, Patch Faster</h3><p>Tesla presents a fascinating contrast. As a software-first car company, Tesla has faced its own security incidents — but its response model has been strikingly different from legacy automakers.</p><p>In 2019 at the Pwn2Own hacking competition, a team from Fluoroacetate broke into a Tesla Model 3 in a way that earned them a $35,000 bounty and the car itself. They exploited a vulnerability in the browser’s renderer to gain control of the infotainment system. Within days, Tesla had an over-the-air patch rolling out to every affected vehicle.</p><p>That turnaround time is essentially impossible for traditional automakers. When your software update process requires physical dealer visits or complex fleet logistics, a days-long patch cycle becomes a months-long exposure window. Tesla’s OTA (over-the-air) update infrastructure — the same technology that adds new features while you sleep — is also a genuine security advantage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ckA-sGJMe5yEuZMSLeATVQ.png" /></figure><p>Tesla also runs a persistent bug bounty program that invites security researchers to find and report vulnerabilities in exchange for cash rewards. This is standard practice in software security, but was essentially unheard of in the auto industry until Tesla normalized it. The logic is simple: you’d rather pay a friendly researcher $10,000 to find a hole than discover it was found by someone with worse intentions.</p><h3>The Attacks You Don’t Read About</h3><p>The Jeep hack makes headlines because it’s dramatic. But there’s a quieter, more common category of automotive attack that affects far more people every year: keyless relay attacks.</p><p>If you have a modern car with keyless entry, here’s something a little unsettling to know. The car and your key fob communicate using short-range radio signals. Your car is constantly listening for that signal, and when it detects the key fob nearby, it unlocks and allows you to start. It seems simple and smart. But “nearby” is determined by signal strength — and signal strength can be manipulated.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FUaYhvV9IvJMQF9g4FqqNQ.png" /></figure><p>The uncomfortable reality is that this attack doesn’t require sophisticated hacking skills or expensive equipment. Relay device kits have been found selling online for a few hundred dollars. And it works against a wide range of vehicles from multiple manufacturers. Insurance fraud investigators and police departments in the UK, US, and Germany have all documented significant increases in keyless theft using this method.</p><p>The fix isn’t straightforward, it requires changing how the authentication protocol works, not just patching a software bug. Some manufacturers have introduced “ultra-wideband” technology that can more accurately determine the key’s physical location, and others have moved to motion sensors that put the key fob to sleep when stationary. But millions of existing vehicles remain vulnerable.</p><h3>Mapping the Threat Landscape</h3><p>When security professionals talk about a “threat landscape,” they mean all the different ways something can be attacked. Modern vehicles have an attack surface that has expanded dramatically over the past decade — and continues to grow with each new connected feature.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QPEljR6HoSSMyvy1SSkfMg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*s3GRz58g00IRreFvjqnGCw.png" /></figure><p>The supply chain threat deserves special attention because it’s the hardest to defend against. A modern vehicle contains software and hardware components from hundreds of suppliers. If any one of those suppliers ships a compromised component — whether through their own negligence or because they were targeted by an adversary — that compromise can propagate into millions of vehicles. This isn’t hypothetical; it’s a known attack pattern in other industries, and automotive is increasingly exposed to it.</p><blockquote><strong>⚠ The OTA Update Paradox: </strong>Over-the-air update capability is simultaneously one of the best security tools in automotive and one of the most dangerous attack surfaces. A compromised OTA system doesn’t just affect one car — it potentially affects every car in a fleet. Securing the update pipeline (cryptographic signing, secure boot, integrity verification) is now one of the most critical tasks in automotive cybersecurity.</blockquote><h3>How Regulators Responded</h3><p>The automotive industry is not short of opinions about how regulation tends to work: slow, reactive, and occasionally painful to implement. But give credit where it’s due, the cybersecurity regulatory response to these threats has been reasonably substantive, even if it took longer than it should have.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cHKssmARiV7VAumLrzjjiQ.png" /></figure><p>Two frameworks dominate the current landscape, and understanding both is essential if you work anywhere in the automotive supply chain.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mAF6XwQma8GcVrMuakWfog.png" /></figure><p>The ISO/SAE 21434 standard is worth understanding beyond the acronym. Its core methodology , TARA requires engineers to systematically identify every asset in a vehicle that could be attacked, analyze the potential threats to those assets, assess the risk of each threat, and design mitigations accordingly. That sounds like common sense, but doing it rigorously across a vehicle with hundreds of ECUs and thousands of software components is genuinely difficult work.</p><p>The standard also does something that matters enormously in practice: it places security obligations on suppliers, not just OEMs. A tier-2 supplier providing a telematics module must now demonstrate cybersecurity engineering — not just functional safety. This has rippled through the entire supply chain and created a significant demand for engineers who understand both domains.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yADoczXci_WOFfWN03FbVw.png" /></figure><h3>Hardware Security: The Silicon Layer</h3><p>Software defenses are essential, but they have an inherent limitation: they run on hardware, and if that hardware is compromised at boot time, no amount of software checking will save you. This is why Hardware Security Modules — HSMs — have become central to automotive cybersecurity architecture.</p><p>An HSM is a dedicated piece of silicon designed to handle cryptographic operations and store secret keys in a way that’s physically resistant to tampering. In an automotive context, an HSM embedded in a vehicle’s main microcontroller does a few critical things. It verifies at boot that the software being loaded is legitimate and hasn’t been tampered with (secure boot). It stores and manages the cryptographic keys used to authenticate communications between ECUs. It provides an isolated execution environment for security-critical operations that malicious software running elsewhere in the system cannot access.</p><blockquote><strong>Why HSMs Matter: </strong>Without hardware-anchored trust, an attacker who compromises the main software environment can fake any security check. With an HSM, even a fully compromised operating system cannot forge signatures or access cryptographic keys stored in the secure enclave. It’s the difference between a lock that only works when nothing is broken and a lock that works even when everything else has failed.</blockquote><p>The challenge is that HSMs add cost and complexity. In an industry that measures margins carefully and designs ECUs to last fifteen years, integrating new silicon is not trivial. The industry has made progress — HSMs are increasingly common in new designs — but the installed base of vehicles without them is enormous and will remain on roads for decades.</p><h3>What This Means for Engineers</h3><blockquote><strong>These aren’t abstract recommendations — they’re the gaps I’ve seen firsthand across OEM and Tier-1 projects, and the skills that separate engineers who can talk security from those who can actually design it.</strong></blockquote><p>If you’re an automotive engineer who trained primarily in mechanical systems, powertrain, or even classical embedded software, this landscape can feel like it arrived overnight. In one sense, it did. The industry went from “cybersecurity is not our problem” to “cybersecurity is a regulatory requirement for market access” in roughly a decade.</p><p>The demand for people who can bridge automotive engineering and cybersecurity is genuinely significant — and the skill set is rare. You need to understand vehicle architecture deeply enough to reason about what matters if compromised, but also understand security engineering well enough to design proper defenses. That combination doesn’t appear automatically in either a traditional automotive engineering program or a computer science security curriculum.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cOSjaTNdfOcAVpSVdlNZfQ.png" /></figure><p>Certifications are beginning to emerge that formalize this knowledge — including programs aligned specifically to ISO/SAE 21434. While no certification replaces hands-on experience, they do help engineers demonstrate competence to employers and provide a structured framework for self-study. If you’re looking to transition into this space, the combination of the 21434 standard itself (which is public) and a structured training program will give you a significant head start.</p><h3>Where This Goes Next</h3><p>Autonomous vehicles add a new dimension that makes everything discussed here higher stakes. A vehicle that drives itself based on sensor data can be attacked not just through its network interfaces, but by manipulating what its sensors perceive — spoofing GPS signals, confusing LIDAR with carefully positioned reflectors, fooling camera-based systems with adversarial patterns. These are active research areas with real demonstrations behind them.</p><p>V2X (Vehicle-to-Everything) communication — the emerging technology that allows vehicles to communicate with each other and with road infrastructure — introduces yet another protocol layer to secure. And as vehicles become more tightly integrated with cloud backends for navigation, insurance telemetry, and feature delivery, the attack surface extends well beyond the vehicle itself into data centers and mobile applications.</p><p>The honest picture is this: the attack surface is expanding faster than the defenses are maturing. That’s not a reason for panic — it’s a reason for urgency. The industry has recognized the problem, regulation is compelling action, and there’s a genuine generation of engineers being built around this discipline. But it remains one of the more serious open challenges in consumer product security today.</p><p>In the next part of this series, we’ll go deeper into threat modeling — specifically, how TARA works in practice, what it looks like to apply it to a real vehicle system, and where teams typically get it wrong. If you’re an engineer starting to work with ISO/SAE 21434, that’s the piece you won’t want to miss.</p><h3>Stay Connected</h3><p>Automotive cybersecurity is evolving rapidly, and new threats continue to emerge as vehicles become more connected and software-defined.</p><p>I will continue publishing deep insights, real-world attack analysis, security architecture breakdowns, and updates on emerging automotive cyber threats and compliance standards. If you’re working in IVI, ECU firmware, SDV platforms, or connected vehicle ecosystems, I invite you to follow this publication and stay informed.</p><p>Your feedback is valuable — feel free to share thoughts, corrections, or discussion points in the comments.</p><p>For professional discussions, collaboration, or technical exchanges, you can reach out directly at:<br> <strong>sharmabudhdi@gmail.com</strong></p><p>Let’s build safer vehicle platforms together.</p><h4>About Me</h4><blockquote>Automotive Software Architect with a decade of experience across OEM and Tier-1 environments. His background spans AOSP architecture, ECU design, embedded firmware, and adversarial AI security in edge and embedded systems. <a href="https://www.linkedin.com/in/budhdisharma/">Connect on LinkedIn →</a></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3da19a30b76e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/securedrive-systems/when-your-car-gets-hacked-the-real-world-of-automotive-cybersecurity-3da19a30b76e">When Your Car Gets Hacked: The Real World of Automotive Cybersecurity</a> was originally published in <a href="https://medium.com/securedrive-systems">SecureDrive Systems</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>