<?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 osint-monitor on Medium]]></title>
        <description><![CDATA[Stories by osint-monitor on Medium]]></description>
        <link>https://medium.com/@info.osintmonitor?source=rss-edfe6a2c181f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*3Z9Jm48nHCcdOBC0</url>
            <title>Stories by osint-monitor on Medium</title>
            <link>https://medium.com/@info.osintmonitor?source=rss-edfe6a2c181f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 01:14:50 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@info.osintmonitor/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[Building ANTARES: Simulating Disinformation Campaigns on a CPU-Only VPS]]></title>
            <link>https://medium.com/@info.osintmonitor/building-antares-simulating-disinformation-campaigns-on-a-cpu-only-vps-3fe611e4f4b8?source=rss-edfe6a2c181f------2</link>
            <guid isPermaLink="false">https://medium.com/p/3fe611e4f4b8</guid>
            <category><![CDATA[disinformation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[fake-news]]></category>
            <category><![CDATA[threat-intelligence]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[osint-monitor]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 20:06:25 GMT</pubDate>
            <atom:updated>2026-11-19T20:27:56.311Z</atom:updated>
            <content:encoded><![CDATA[<p>Most disinformation doesn’t look like obvious “fake news” anymore. It looks like a slow, coordinated shift in how people frame an event, distrust a source, or pick a side. If you work in defense or threat intelligence, reading about campaigns after the fact only gets you so far. You also want to see how a narrative gets bent from the inside.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QdqdA4nhI1ynXN0_" /></figure><p>That’s why I started building <strong>ANTARES</strong> (<em>Adversarial Narrative Threat Analysis and Response Evaluation System</em>). The idea is simple: simulate how hostile actors twist a piece of news, and immediately pair that simulation with defensive response material. No offense, no publishing propaganda. Just stress-testing narratives so defenders aren’t caught flat-footed.</p><p>There was one catch from day one: I wanted to run it on my cheap VPS with <strong>4 vCPUs, 8 GB of RAM, and no GPU</strong>. Here is how the system is put together and what I had to work around to keep it alive on constrained hardware.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hIuNnOVlRoTGatGuRVSy2A.png" /><figcaption>ANTARES homepage</figcaption></figure><p>ANTARES supports two backends so I can switch depending on what I’m testing and how sensitive the input is:</p><ul><li><strong>Google Gemini (cloud)</strong>: faster, handles heavier reasoning, subject to standard API policies.</li><li><strong>Local models via </strong><strong>llama.cpp</strong>: stays on the box, no external calls, no content filters getting in the way of red-team phrasing.</li></ul><p>Running a local model on CPU changes how you write prompts. Long, heavily structured prompts that work fine on Gemini will eat time and context on a small CPU setup. For llama.cpp, I had to strip the prompts down to the bone, just to keep runtimes reasonable and avoid choking the machine.</p><p>A simulation isn’t instantaneous, especially on CPU. If the web app waited on the model inline, requests would time out and the UI would feel broken. Instead, tasks go into a <strong>SQLite-backed queue</strong>, and a <strong>Python worker running under </strong><strong>systemd</strong> pulls jobs one by one in the background. You can queue up multiple runs, watch status updates on a dashboard, and let the server chew through them at its own pace.</p><h3>How a Simulation Runs</h3><p>The pipeline follows a straightforward flow from raw text to exportable intelligence.</p><h3>1. Language and Cultural Framing</h3><p>You pick <strong>English or Italian</strong> up front. That changes more than labels and buttons: the prompts shift to account for different media ecosystems, cultural reference points, and geopolitical fault lines in Anglo-Saxon vs. Italian contexts.</p><h3>2. Text Ingestion and Actor Matching</h3><p>You paste a starting text (a press release, an article, a public statement) and pick a threat actor from a local library (for example, <em>APT28</em>, <em>Doppelganger</em>, or <em>CyberBerkut</em>).</p><p>Each actor profile is a JSON file mapping known tactics, narrative habits, communication channels, and cognitive targets, tied back to <strong>MITRE ATT&amp;CK</strong> and <strong>DISARM</strong>. To save time scrolling through profiles, I added a <strong>semantic embedding matcher</strong> that compares the input text against actor profiles and suggests the closest fit.</p><h3>3. Red Team, Then Blue Team</h3><p>Once the worker picks up a job, it injects the selected actor’s JSON profile and runs two passes in sequence:</p><ul><li><strong>Red Team pass</strong>: models how that actor would frame or distort the source text (likely angles, target audiences, sample messaging hooks, and risk notes) returned as structured JSON.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WyuPbh1I2cosHZllBLqcEA.png" /></figure><ul><li><strong>Blue Team pass</strong>: takes that simulated attack and drafts defensive material (counter-narratives, messaging guidance, debunking angles, and escalation notes).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n3KWljcvB2-Ojg-j_mq_Jg.png" /></figure><h3>4. Exports for Existing Workflows</h3><p>An analysis isn’t much use if it stays trapped in a web view. Finished runs can be exported as:</p><ul><li><strong>PDF</strong>, for standard reading and sharing.</li><li><strong>STIX 2.1</strong>, for structured threat intelligence exchange.</li><li><strong>MISP</strong>, formatted as an event with attributes ready to drop into security tooling.</li></ul><h3>Comparing Actors Side by Side</h3><p>One of the more useful parts of the app is the <strong>multi-actor comparison view</strong>. A state-backed group and an opportunistic hacktivist group won’t spin the same story the same way: their incentives, tone, and targets differ.</p><p>You can take a completed analysis, clone it across up to three additional actors, and view the resulting attack-and-defense outputs side by side in a matrix, drill into individual runs, or export the comparison to <strong>PDF or CSV</strong>.</p><h3>Responsible Use</h3><p>A quick line on scope: ANTARES is meant for defensive analysis and training, not generating content to publish. Every simulated attack output is paired with defensive countermeasures, and the actor profiles rely on publicly documented TTPs from frameworks like MITRE ATT&amp;CK and DISARM. If you fork it or adapt it, keep it grounded in defensive research.</p><h3>Demo and Source Code</h3><p>Because the server is small and I don’t want anonymous traffic hammering CPU inference, the public deployment runs in <strong>Demo Mode</strong>. Anyone can browse existing reports, inspect actor profiles, and explore the comparison matrix, while triggering new simulations requires admin auth.</p><p>👉 <a href="https://osintmonitor.ddns.net/ANTARES/">https://osintmonitor.ddns.net/ANTARES/</a></p><p>The code is open source if you want to dig into the worker loop, the SQLite queue, or the split prompt setup:</p><p>👉 <a href="https://github.com/osint-monitor/ANTARES">https://github.com/osint-monitor/ANTARES</a></p><p>If you’ve built similar pipelines on tight hardware or use adversarial framing in your own threat intel work, I’d be curious how you handle it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3fe611e4f4b8" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>