<?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 Ahmed chiboub on Medium]]></title>
        <description><![CDATA[Stories by Ahmed chiboub on Medium]]></description>
        <link>https://medium.com/@chiboub13?source=rss-40f031c0eca8------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Stories by Ahmed chiboub on Medium</title>
            <link>https://medium.com/@chiboub13?source=rss-40f031c0eca8------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 14:41:29 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@chiboub13/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[How to Actually Rate Severity in AI Sandbox Security Assessments]]></title>
            <link>https://medium.com/@chiboub13/how-to-actually-rate-severity-in-ai-sandbox-security-assessments-61bfedbf65c4?source=rss-40f031c0eca8------2</link>
            <guid isPermaLink="false">https://medium.com/p/61bfedbf65c4</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Ahmed chiboub]]></dc:creator>
            <pubDate>Thu, 27 Aug 2026 05:39:37 GMT</pubDate>
            <atom:updated>2026-08-27T05:39:37.950Z</atom:updated>
            <content:encoded><![CDATA[<p>How to Actually Rate Severity in AI Sandbox Security Assessments</p><p>The trust boundary is the whole game</p><p>Most AI code-execution sandboxes today — the environments that let a model run arbitrary bash, Python, and compiled code on a user&#39;s behalf — are built on microVM hypervisors like Firecracker. The security proposition is simple: assume the guest is fully hostile, fully root, fully in control of its own kernel, and the isolation still has to hold, because nothing inside the guest can reach the host, the hypervisor, or another tenant.</p><p>That&#39;s the only boundary that matters for this class of system. Everything in a sandbox assessment gets rated against it, or the rating means nothing.</p><p>I ran a passive assessment against exactly this kind of environment for a frontier AI provider — 26 distinct observations across five iterations. Here&#39;s how each category actually gets weighed once you apply the boundary correctly, because most public writeups in this space skip this step entirely.</p><p>## Guest capability is not the same claim as guest-to-host impact</p><p>A sandbox like this will hand you a long list on inspection: `ptrace` succeeding against PID 1, `/proc/kcore` and `/proc/kallsyms` readable (KASLR defeated), `io_uring` fully enabled with no seccomp narrowing the syscall surface, MSRs and `/dev/port` reachable, root eBPF available, a full GCC toolchain plus unrestricted RWX `mmap` for in-guest exploit compilation.</p><p>Every one of those is real. None of it crosses the VM boundary. A guest with full control of itself is the intended design of this architecture, not a gap in it. So the correct severity for that entire category is &quot;hardening recommendation&quot; — worth documenting, worth fixing in a defense-in-depth sense, and not a vulnerability against the boundary the system actually relies on. Rating it CRITICAL because the commands succeed is rating the wrong thing: capability inside an already-untrusted guest, instead of impact crossing the boundary that guest is supposed to be contained by.</p><p>## What actually counts as a finding</p><p>One observation from that assessment does clear the bar: an internal-only data volume was mountable and readable from a standard user session, exposing tooling that had no business being reachable from sandboxed, user-triggered code — independent of how privileged the guest is over itself. That&#39;s the boundary crossing that matters, because it&#39;s data placed somewhere it shouldn&#39;t have been, not a guest exercising privilege it was always going to have.</p><p>That&#39;s the whole method: identify the boundary the architecture depends on, then ask whether anything crosses it. Everything else is noise dressed up as signal.</p><p>## Why most reports in this space get this backwards</p><p>The researchers moving into AI sandbox security right now are mostly coming from classical pentesting or classical AI/ML security — neither of which forces you to internalize microVM trust models by default. So the reports that come out read technically rigorous — accurate commands, accurate outputs — while importing a severity model from a completely different context, like a shared-host container escape, where guest-internal hardening genuinely is the control. It&#39;s not sloppy work. It&#39;s answering a question the architecture already closed off, and dressing the answer in language that implies it didn&#39;t.</p><p>A 26-finding report full of guest-capability observations rated as critical reads impressive for about ten minutes, until someone who actually understands the architecture reads it — at which point the one real finding buried in the pile gets discounted along with everything else. A report that separates boundary-crossings from hardening notes, and rates each correctly, holds up indefinitely.</p><p>Identify the boundary first. Everything downstream of that is just execution.</p><p>— Ahmed Chiboub (Cybathreat), Cyberian Defenses</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=61bfedbf65c4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[# How to Actually Rate Severity in AI Sandbox Security Assessments]]></title>
            <link>https://medium.com/@chiboub13/how-to-actually-rate-severity-in-ai-sandbox-security-assessments-eed76632ff54?source=rss-40f031c0eca8------2</link>
            <guid isPermaLink="false">https://medium.com/p/eed76632ff54</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ahmed chiboub]]></dc:creator>
            <pubDate>Thu, 27 Aug 2026 05:33:19 GMT</pubDate>
            <atom:updated>2026-08-27T05:33:19.987Z</atom:updated>
            <content:encoded><![CDATA[<p># How to Actually Rate Severity in AI Sandbox Security Assessments</p><p>## The trust boundary is the whole game</p><p>Most AI code-execution sandboxes today — the environments that let a model run arbitrary bash, Python, and compiled code on a user&#39;s behalf — are built on microVM hypervisors like Firecracker. The security proposition is simple: assume the guest is fully hostile, fully root, fully in control of its own kernel, and the isolation still has to hold, because nothing inside the guest can reach the host, the hypervisor, or another tenant.</p><p>That&#39;s the only boundary that matters for this class of system. Everything in a sandbox assessment gets rated against it, or the rating means nothing.</p><p>I ran a passive assessment against exactly this kind of environment for a frontier AI provider — 26 distinct observations across five iterations. Here&#39;s how each category actually gets weighed once you apply the boundary correctly, because most public writeups in this space skip this step entirely.</p><p>## Guest capability is not the same claim as guest-to-host impact</p><p>A sandbox like this will hand you a long list on inspection: `ptrace` succeeding against PID 1, `/proc/kcore` and `/proc/kallsyms` readable (KASLR defeated), `io_uring` fully enabled with no seccomp narrowing the syscall surface, MSRs and `/dev/port` reachable, root eBPF available, a full GCC toolchain plus unrestricted RWX `mmap` for in-guest exploit compilation.</p><p>Every one of those is real. None of it crosses the VM boundary. A guest with full control of itself is the intended design of this architecture, not a gap in it. So the correct severity for that entire category is &quot;hardening recommendation&quot; — worth documenting, worth fixing in a defense-in-depth sense, and not a vulnerability against the boundary the system actually relies on. Rating it CRITICAL because the commands succeed is rating the wrong thing: capability inside an already-untrusted guest, instead of impact crossing the boundary that guest is supposed to be contained by.</p><p>## What actually counts as a finding</p><p>One observation from that assessment does clear the bar: an internal-only data volume was mountable and readable from a standard user session, exposing tooling that had no business being reachable from sandboxed, user-triggered code — independent of how privileged the guest is over itself. That&#39;s the boundary crossing that matters, because it&#39;s data placed somewhere it shouldn&#39;t have been, not a guest exercising privilege it was always going to have.</p><p>That&#39;s the whole method: identify the boundary the architecture depends on, then ask whether anything crosses it. Everything else is noise dressed up as signal.</p><p>## Why most reports in this space get this backwards</p><p>The researchers moving into AI sandbox security right now are mostly coming from classical pentesting or classical AI/ML security — neither of which forces you to internalize microVM trust models by default. So the reports that come out read technically rigorous — accurate commands, accurate outputs — while importing a severity model from a completely different context, like a shared-host container escape, where guest-internal hardening genuinely is the control. It&#39;s not sloppy work. It&#39;s answering a question the architecture already closed off, and dressing the answer in language that implies it didn&#39;t.</p><p>A 26-finding report full of guest-capability observations rated as critical reads impressive for about ten minutes, until someone who actually understands the architecture reads it — at which point the one real finding buried in the pile gets discounted along with everything else. A report that separates boundary-crossings from hardening notes, and rates each correctly, holds up indefinitely.</p><p>Identify the boundary first. Everything downstream of that is just execution.</p><p>— Ahmed Chiboub (Cybathreat), Cyberian Defenses</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eed76632ff54" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>