<?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[Seenu on Agentic AI - Medium]]></title>
        <description><![CDATA[Notes on building grounded, secure agentic AI, from a senior security engineer. How these systems work, how to keep them honest, and how to harden them, without the hype. - Medium]]></description>
        <link>https://medium.com/seenu-on-agentic-ai?source=rss----9fdefa10c7f9---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Seenu on Agentic AI - Medium</title>
            <link>https://medium.com/seenu-on-agentic-ai?source=rss----9fdefa10c7f9---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 05:18:58 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/seenu-on-agentic-ai" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Building a grounded, secure AI assistant: the why, and the honest how]]></title>
            <link>https://medium.com/seenu-on-agentic-ai/building-a-grounded-secure-ai-assistant-the-why-and-the-honest-how-a1c4fb1b2058?source=rss----9fdefa10c7f9---4</link>
            <guid isPermaLink="false">https://medium.com/p/a1c4fb1b2058</guid>
            <dc:creator><![CDATA[Seenu Bhupala]]></dc:creator>
            <pubDate>Sat, 04 Jul 2026 06:49:35 GMT</pubDate>
            <atom:updated>2026-10-17T01:29:01.959Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7DyeQY0d_yfiKD_UXSbCPw.png" /><figcaption>Defender for Servers Expert. My grounded, secure agentic AI assistant.</figcaption></figure><h4><strong><em>A personal project, and the security thinking behind it.</em></strong></h4><p><strong>Try it live</strong>: <a href="https://ai.cyberfixpro.com/">ai.cyberfixpro.com</a></p><p>I have spent a fair bit of time watching AI assistants answer with complete confidence and, often, complete fiction. As a security engineer, that sits badly with me. A confident wrong answer is not a small thing. In security work it can send someone down the wrong path, and it quietly erodes trust in the tool, and in the person who stood it up.</p><p>I will keep the deep mechanics to myself. Not to be precious about it, but because the value in this kind of work is in the judgement, not the recipe. What follows is the shape of the approach, which is the part actually worth sharing.</p><p><strong>Why I built it</strong></p><p>I learn best by building the whole thing, end to end, and breaking it a few times along the way. I wanted to understand agentic AI properly, not as a demo but as something that had to stand up in the open, on my own domain.</p><p>I also wanted to see whether my day job instincts carried across. I spend my time onboarding customers to security platforms and thinking about how systems fail. An AI assistant is just another system that can fail, so I treated it like one.</p><p><strong>What it actually is</strong></p><p>It is a narrow, grounded assistant. It answers questions within one defined subject area and politely declines everything else. Every factual claim it makes is tied back to a real, current source, with a link you can click and verify. If it cannot find grounding for something, it says so rather than filling the gap with something that sounds right.</p><p>That last behaviour is the whole point. Most of the trouble with these systems comes from their willingness to be helpful past the limit of what they actually know.</p><p><strong>Grounding was the entire idea</strong></p><p>There are two ways an assistant can answer. It can reach into what the model absorbed during training, which is broad but frozen in time and impossible to verify, or it can retrieve from a trusted, current source and answer only from what it just read. The first feels magical and occasionally invents things with great conviction. The second is slower to build and far more honest.</p><p>I built for the second. The assistant retrieves first, reasons over what it found, and then answers with citations attached. The rule underneath it is simple to say and surprisingly hard to enforce: if you cannot cite it, do not assert it.</p><p>The failure mode I was most worried about is the well dressed fabrication, where the model produces a tidy, structured, confident answer with a citation that looks perfect and points at nothing real. That one is dangerous precisely because it looks correct. A good chunk of the work went into making that outcome difficult.</p><p><strong>How it works, from a distance</strong></p><p>Without going into the plumbing, there are a handful of moving parts. A language model does the reasoning. A retrieval layer gives it access to trusted, current material rather than relying on memory. A tool connection lets it reach official documentation live, so its answers track the source of truth instead of a snapshot. An orchestration layer, the agent itself, ties the instructions, the tools and the knowledge together and decides what to do with a given question. A small web front end sits on top so people can actually use it, on a real domain.</p><p>The order matters more than the pieces. Retrieve, then reason, then answer, then cite. Get that sequence right and a lot of the bad behaviour simply has fewer places to hide.</p><p><strong>The infrastructure, at a glance</strong></p><p>I kept it deliberately lean and cloud hosted. There is a managed service running the model and the agent, a retrieval and knowledge layer feeding it grounded material, a lightweight web application, and a domain with an edge in front of it. Nothing exotic.</p><p>One habit from the day job carried straight over. Once something was no longer earning its keep, I removed it. At one point I realised part of the grounding setup was costing me money for capability the design no longer needed, so I decommissioned it cleanly and the assistant carried on without missing a beat. Knowing what to take out is as much a part of architecture as knowing what to put in.</p><p><strong>Securing it, which is really the point</strong></p><p>I approached the assistant the way I would approach any system that will be probed by strangers. Assume hostile input, assume someone will try to make it misbehave, and design so that no single failure is catastrophic. Defence in depth, in other words, applied to a slightly unusual system.</p><p>A few of the layers, described in principle rather than in detail:</p><p>Scope, as least privilege for topics. It answers what it is meant to and declines the rest. A narrow, well defined remit is a security control, not just a product decision, because it removes a huge amount of surface area.</p><p>Grounding as a safety mechanism. Because it will not assert without a source, an entire class of confident nonsense is closed off by design rather than patched after the fact.</p><p>Resistance to prompt injection and jailbreaks. All incoming text, and all retrieved text, is treated as untrusted data rather than as instructions. Attempts to smuggle in new orders, to trigger a persona change, or to talk it into ignoring its rules are recognised and refused, quietly and consistently.</p><p>Confidentiality of its own workings. It will not recite its instructions or describe its internals, no matter how the question is framed. Reconnaissance is the first step of most attacks, so I gave it nothing to hand over.</p><p>Output discipline. It answers in plain, readable language and will not be coaxed into odd encodings that could be used to sneak content past a filter.</p><p>Content safety at the platform edge, paired with a single, uniform refusal. Whichever layer catches a bad request, the person on the other end gets the same calm, consistent response. The experience should never leak which defence just fired.</p><p>Citation integrity. It cites only real, verbatim source links. It does not tidy up, shorten or reconstruct a link from memory, because a link that looks right but leads nowhere is its own small breach of trust.</p><p>Evaluation and regression testing. I keep a fixed set of questions, including deliberately adversarial ones, and re-run the lot after every change. If a future tweak quietly weakens a guardrail, that battery is meant to catch it before anyone else does.</p><p>None of these is clever on its own. Together they mean that getting the assistant to do something it should not requires defeating several independent controls at once, which is exactly the position you want an attacker to be in.</p><p><strong>Choosing a model, and knowing when to change it</strong></p><p>I started on a small, inexpensive model and it mostly worked. What it did not do reliably was hold a nuanced boundary. It would occasionally refuse a perfectly valid question, then, on another run, let through something it should have declined. That inconsistency is not just annoying, it is a reliability and security problem, because a guardrail that only works most of the time is not really a guardrail.</p><p>Moving to a stronger model fixed the wobble. It followed the same instructions far more consistently, in both directions. The lesson I took from it is that model choice is an engineering and security decision, not only a quality one.</p><p>If you take one thing from this, take the mindset rather than the settings. Ground everything. Assume the input is hostile. Test relentlessly. The specifics will follow from those three.</p><p><strong>Where it is now</strong></p><p>It is live, and it holds up. It answers within its remit, grounded and cited, and it stays composed when people try to break it. It is a personal proof of concept, not a product, and that is exactly what I wanted. It let me learn the whole stack, from the model down to the last guardrail, and it gave my security instincts somewhere new to be useful.</p><p>Much of what I know came from people who shared their thinking openly, and asked for nothing in return. This is a small attempt to do the same, minus the parts that are better learnt by doing.</p><p>A longer, more detailed write up may follow. For now, if you are building in this space, I hope the shape of it is useful.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a1c4fb1b2058" width="1" height="1" alt=""><hr><p><a href="https://medium.com/seenu-on-agentic-ai/building-a-grounded-secure-ai-assistant-the-why-and-the-honest-how-a1c4fb1b2058">Building a grounded, secure AI assistant: the why, and the honest how</a> was originally published in <a href="https://medium.com/seenu-on-agentic-ai">Seenu on Agentic AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>