<?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 Abdullah Mohammed on Medium]]></title>
        <description><![CDATA[Stories by Abdullah Mohammed on Medium]]></description>
        <link>https://medium.com/@ArkhamHunter?source=rss-f52da4524c31------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*rKsq3rBnUhBRLYIWTzbOhQ.jpeg</url>
            <title>Stories by Abdullah Mohammed on Medium</title>
            <link>https://medium.com/@ArkhamHunter?source=rss-f52da4524c31------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 07:29:06 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ArkhamHunter/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[NTFS Analysis]]></title>
            <link>https://medium.com/@ArkhamHunter/ntfs-analysis-17252a39ee5d?source=rss-f52da4524c31------2</link>
            <guid isPermaLink="false">https://medium.com/p/17252a39ee5d</guid>
            <dc:creator><![CDATA[Abdullah Mohammed]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 11:20:38 GMT</pubDate>
            <atom:updated>2026-11-18T11:20:38.803Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>NTFS (New Technology File System) is the default file system for Windows operating systems, developed by Microsoft. This file system is known for its robustness, scalability, advanced features like file permissions, encryption, journaling, and support for large files and partitions. NTFS organizes data in a way that makes it efficient for the operating system and significant for forensic investigations.</p><h3>NTFS OVERVIEW</h3><p>As already covered, NTFS, or New Technology File System, is a file system developed by Microsoft. It’s been the default file system for Windows operating systems since Windows NT 3.1, way back in 1993. A file system is the way our computer organizes and stores files on a drive.</p><p>It is trivial to understand that NTFS can also be classified as a Journaling file system because it is designed to maintain the file system’s consistency by recording or writing down the changes within the file system in a journal (log) before applying them to the main file system structures. From an OS point of view, this is useful when recovering the OS from unexpected system crashes, power failures, etc. NTFS has some features that make it stand out from other file systems.</p><p>Some of the key features of NTFS are :</p><ul><li><strong>Advanced Metadata : NTFS maintains detailed metadata, including file creation, modification, and access times, which is invaluable for forensic investigations.</strong></li><li><strong>Journaling : NTFS uses a journaling feature to keep track of changes. This feature helps recover data after crashes and provides a trail of file system activities. This can also be helpful for forensics investigations</strong></li><li><strong>Security Features : NTFS supports advanced security features like encryption (EFS) and permissions, which can be analyzed to understand access patterns and security breaches.</strong></li><li><strong>Large Volume Support :</strong> NTFS can handle large volumes and files, making it suitable for modern storage needs</li><li><strong>Resilience : NTFS is more resilient to corruption compared to older file systems like FAT32, thanks to its robust structure and recovery mechanisms</strong></li><li><strong>Compatibility :</strong> While primarily for Windows, NTFS can be read (and sometimes written) on Linux and macOS with third-party tools</li><li><strong>Compression : It includes built-in file compression to save space</strong></li></ul><p>NTFS Components</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/741/1*Zm3LQbxhkxvU9RxlOcyB4Q.png" /></figure><h4>Partition Boot Sector (PBS)</h4><p>The Partition Boot Sector is the first sector of an NTFS volume and plays a vital role in the booting process. It contains the necessary information to locate and load the operating system.</p><p>It also includes the BIOS Parameter Block (BPB), which defines the disk layout. It maintains critical file system parameters such as sectors per cluster and the location of the Master File Table</p><p><strong>Key Contents</strong></p><p>Some of the key components of PBS are:</p><ul><li>Jump instruction to bootstrap code.</li><li>File system type indicator (e.g., NTFS).</li><li>Location of the MFT and the MFT Mirror.</li><li>End-of-sector marker (<strong>0x55AA</strong>).</li></ul><p><strong>Forensics Value</strong></p><p>Looking at the PBS from the forensics perspective can reveal information about the boot process and disk configuration. This sector can be analyzed to determine whether it has been tampered with (e.g., by malware or rootkits) or to validate the integrity of the boot code and the BIOS Parameter Block (BPB)</p><h4>Master File Table (MFT)</h4><p>One of the key components/sectors of the NTFS disk is the Master File Table (MFT). This is considered the backbone of NTFS, acting as a database that contains metadata about every file and directory. It provides a comprehensive structure for the local file data and can be used to track all the files and directories on the disk, including the user data and the system files.</p><p><strong>Characteristics</strong></p><p>Some of the key characteristics of MFT are:</p><ul><li>Divided into fixed-size records (typically 1 KB).</li><li>Each file and directory gets an MFT entry.</li><li>File attributes such as timestamps, permissions, and data location are stored as metadata.</li></ul><p><strong>Forensics Value</strong></p><p>the deleted files records may still exist in the MFT , and its essential for timeline analysis , as it contains the creation , modification , access time</p><p><strong>$MFTMirr</strong></p><p>you could say its semi-backup since it duplicates the first few records from the MFT so we ensure redundancy</p><p><strong>Forensics Value</strong></p><p>From a forensics point of view, it can be used to cross-verify the integrity of the MFT. If the primary MFT is damaged or tempered, it can be used to get useful metadata.</p><p><strong>System Files</strong></p><ul><li>$MFT: Master File Table.</li><li>$MFTMirr: MFT Mirror.</li><li>$LogFile: A transactional log for consistency.</li><li>$Bitmap: Tracks allocated and free clusters.</li><li>$Boot: Stores the boot sector information.</li><li>$BadClus: Tracks bad sectors.</li><li>$UpCase: Stores uppercase character mapping for case-insensitive comparisons.</li></ul><h4>File Data Area</h4><p>the area in which the user and system files is stored , either within the mft recored itself (resident — if the file content is small enough ) or it stores the actual file content of files and directories within cluster ( non-resident)</p><h4>Alternate Data Stream</h4><p>its a feature of NTFS that allows you store multiple data streams into a single file or directory , while the primary data stream contains the main content of a file , ADS can store additional metadata or hidden content without altering the size or appearance of the main file</p><h3>MFT Record Analysis</h3><p>he Master File Table (MFT) Record is considered the most important file for the forensics investigator because the MFT handles any file created/modified/deleted on the file system. The Windows OS creates an entry for each individual file or folder within the MFT record.</p><h4>Master File Table</h4><p>The MFT record is a core component of the NTFS, which serves as the database that tracks all files and directories. Each file or directory on the disk has a corresponding MFT record, which acts as a detailed metadata repository for that object.</p><p>The MFT record contains the following information:</p><ul><li><strong>File Name:</strong> The file’s name or the directory contained within.</li><li><strong>Standard Information</strong>: Contains key information about the link count, timestamps, file attributes, etc.</li><li><strong>Attributes:</strong> Metadata like MACB timestamp, attributes like Read-only, hidden, etc.</li><li><strong>File Data:</strong> This contains the actual data if the file is small enough to fit within the MFT record; otherwise, it points to the clusters where the data is stored.</li><li><strong>File Index:</strong> Index information that points to other files or directories.</li><li><strong>Security Information</strong>: ACL that defines the security permissions for the file or directory.</li></ul><p><strong>Forensics Value</strong></p><p>MFT records play a vital role in investigations, as they are considered an invaluable source of evidence because they provide a comprehensive map of the file system activity and the timeline. Some of the key aspects of MFT records, from the forensics point of view, are:</p><ul><li>It provides evidence of the presence of a certain file on the disk, even if the file is deleted.</li><li>It can also provide evidence of file modification, which could be useful in certain forensics use cases.</li></ul><h3>Examining the MFT Record</h3><p>Now that we have understood the importance of the MFT record, it is time to extract the juicy information from the MFT File we extracted in the previous task and validate all the key findings that we discussed earlier.</p><p>To extract the information from the MFT records, we will use the MFTECmd.exe tool, which is placed in the EZ Tools directory on the Desktop. Open the terminal and run the following command to check the available options, as shown below:</p><p><strong>Command:</strong> MFTECmd.exe --help</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/994/1*HIxTY6fHIfNxJEaNYNvkTw.png" /></figure><p>As mentioned in the example commands, we will be using the following three flags to create our query:</p><ul><li>f: To provide the MFT File Record</li><li>-csv: Directory to save the CSV formatted result to.</li><li>-csvf: File name to save CSV formatted result to.</li></ul><p>Let’s use the flags mentioned above to extract the records from the MFT File present in the ~\Desktop\Evidence\ and save it in the same folder.</p><p><strong>Command:</strong> MFTECmd.exe -f ..\Evidence\$MFT --csv ..\Evidence --csvf ..\Evidence\MFT_record.csv</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*qEmFPZGBO7z8Zq4gbo54zg.png" /></figure><p>If we look at the output, we can see that the tool has correctly identified the file type as MFT. Also, the output is saved in the Evidence folder as</p><p>MFT_record.csv</p><p>. We can click on the output file, and it will automatically load the file into another useful tool called Timeline Explorer, as shown below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hpXj8Gl9x4R84YbtanJxAA.png" /></figure><h3>Important Columns in the MFT Record</h3><p>If we examine the output, we can see that various columns contain important information. Let’s break down these columns to better understand the output.</p><ul><li><strong>Entry Number</strong></li><li><strong>Description:</strong> A unique identifier for the MFT record.</li><li><strong>Importance:</strong> Helps map the record to its specific file or directory and reuse track records.</li><li><strong>Parent Entry Number</strong></li><li><strong>Description:</strong> The entry number of the parent directory.</li><li><strong>Importance:</strong> Shows the file or directory location in the file system hierarchy, aiding in reconstructing directory structures.</li><li><strong>Sequence Number</strong></li><li><strong>Description:</strong> A counter incremented when the MFT record is reused.</li><li><strong>Importance:</strong> Helps identify reused records and differentiate between old and new files occupying the same entry.</li><li><strong>File Name</strong></li><li><strong>Description:</strong> The name of the file or directory.</li><li><strong>Importance:</strong> Essential for identifying the file and matching it with user activities or processes.</li><li><strong>Timestamps</strong></li><li><strong>Creation Time:</strong> When the file or directory was created.</li><li><strong>Modification Time:</strong> When the content of the file was last modified.</li><li><strong>Access Time:</strong> When the file was last accessed.</li><li><strong>MFT record Modification Time:</strong> When the metadata was last updated.</li><li><strong>Importance:</strong> Crucial for <strong>timeline analysis</strong>, helping to reconstruct the sequence of events involving the file.</li><li><strong>Flags</strong></li><li><strong>Description:</strong> Indicates whether the record represents a file, directory, or unused record.</li><li><strong>Importance:</strong> Determines the type of object and its activity status (e.g., deleted or active).</li><li><strong>Entry Flags</strong></li><li><strong>Description:</strong> Flags like Read-only, Hidden, System, etc.</li><li><strong>Importance:</strong> Useful for identifying hidden or system files often used by malware.</li><li><strong>In Use</strong></li><li><strong>Description</strong>: Indicates whether the current record is associated with the active file or directory.</li><li><strong>Importance</strong>: It can indicate which files are present on the disk and which are deleted or no longer in use. MFT File still maintains the record even after the deletion.</li><li><strong>Logical Size</strong></li><li><strong>Description:</strong> The size of the file’s actual data.</li><li><strong>Importance:</strong> Provides insights into the file’s content size, useful for verifying anomalies (e.g., files marked empty but containing data).</li><li><strong>Physical Size</strong></li><li><strong>Description:</strong> The amount of disk space allocated to the file.</li><li><strong>Importance:</strong> Helps identify or fragmented files.</li><li>slack space</li></ul><p>Now that we understand what these columns mean and what key values they provide, explore the MFT record and see if you can find footprints of any suspicious file or the tool.</p><h3>MACB Time</h3><p>In the records, we see different timestamps associated with each record. What are they, and how are they different from each other? Well, they are classified as MACB time. <strong>MACB</strong> refers to the timestamps associated with file system metadata, which provide crucial information about a file’s history. These timestamps are:</p><ul><li><strong>M — Modified</strong>: The time when the file’s content was last changed.</li><li><strong>A — Accessed</strong>: The time when the file was last read or opened.</li><li><strong>C — Changed (Metadata)</strong>: The time when the file’s metadata (e.g., permissions, ownership) was last altered.</li><li><strong>B — Birth (Creation)</strong>: The time when the file was first created (available in NTFS and other advanced file systems).</li></ul><h3>NTFS Journaling</h3><p>NTFS includes a very important feature called file system journaling. This technology allows the OS to maintain a transactional record of all changes made to a volume, so in the event of a crash or power failure, the system can roll back the changes or continue where it left off.</p><p>The goal is to maintain file system integrity and to prevent catastrophic events from occurring. Some of the key features of NTFS Journaling are:</p><ul><li><strong>File system Integrity</strong>: NTFS journaling ensures that the file system can recover even after crashes, minimizing data loss and corruption risks. This is essential for both performance and reliability.</li><li><strong>Deleted Files</strong>: The ability to track when files were deleted through the allows forensic experts to establish timelines critical in investigations.</li><li>USN journal</li><li><strong>USN vs. Log File</strong>: Understanding the difference between the USN journal (file changes) and the log file ( metadata changes) is vital for accurate forensic analysis, as they provide different insights.</li><li>MFT</li><li><strong>Historical Data Access</strong>: Volume shadow copies provide a way to retrieve older versions of journals, allowing forensic teams to analyze data from different points in time, which might reveal crucial evidence.</li></ul><h3>Types Of NTFS Journals</h3><p>There are two journals in the NTFS which we will explore:</p><p><strong>$LogFile</strong></p><p>this file is a special metadata file in the NTFS that acts as the Journal , it Records Metadata changes such as file creation , deletion , modification before committing them on the disk</p><p>If a system failure occurs, NTFS can replay the transactions in $LogFile during the recovery process to restore consistency</p><p>This log file is located in the root directory</p><p><strong>Universal Sequence Number (USN) Journal ($USNJrnl)</strong></p><p>its a feature in NTFS that records changes made to files , directories and their attributes</p><p>It provides a historical record of file system activity, especially useful for monitoring, backup, and indexing services.</p><p>This is located in the $Extend Directory within the root directory</p><p>Structure of USNJrnl</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WtbWQ_aPpO5q6iu57Gbi-Q.png" /></figure><ul><li><strong>$Max</strong>: Defines the maximum size of the journal and its allocation policy.</li><li><strong>$J</strong>: Stores the actual change records. $J is the actual file we are interested in investigating.</li><li>$J, in fact, was implemented as an ADS (Alternative data stream) in NTFS.</li></ul><p><strong>Extract the $J File</strong></p><p>Let’s extract the $J file from the FTK Imager, as shown below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vDLkWUcJFRQ_diNpLsBnlg.png" /></figure><p>MFTECmd.exe -f ..\Evidence\$J — csv ..\Evidence — csvf USNJrnl.csv</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/877/1*ZGCZqXjm4wkRHmMYpmd6uQ.png" /></figure><p>The above command will parse the $J file content and store it in the Evidence folder under the name USNJrnl.csv. It is important to note that this tool correctly identifies the type of input file as the USNJournal, as shown in the image above.</p><p>Now, go to the Evidence folder and open this file. This file will automatically be opened in Timeline Explorer, as shown below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ifo6aF7PJMeMFv0TjK8MDw.png" /></figure><h3>Examining the Updated Reason</h3><p>If we click on the Update Reason column, we can see various reasons, indicating the type of the file system change. Some of the common Opcodes associated with the file changes are explained below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/317/1*t_fy2dCNOitvyRLlmpOUUw.png" /></figure><h3>Index Allocation Attributes ($I30) Overview</h3><p>The $I30 file is an NTFS index attribute, which can also be classified as the directory index that maintains a structured layout of files and directories within an NTFS volume. The Windows OS uses it to store and retrieve information about directory entries. As we examined the directories in FTK Imager, we observed that $I30 is present in some directories containing information about the files and folders within the directory</p><p>Simply put, it is mainly used to store a directory’s index entries and contains metadata about the files and subdirectories within it. This metadata includes filenames, timestamps, and other details about files, even those that have been deleted or modified.</p><h3>Slack Space</h3><p>Slack space in $I30 refers to files that are deleted, renamed, or moved within a directory. Even though a file entry is removed from the active index, residual data may still be present in the slack space, which may be helpful for forensics investigations when looking for deleted files.</p><h3>Forensics Value</h3><p>Now that we understand the $I30 file and its purpose, it’s important to understand its forensics value. Some of the key forensics values we can get from this are:</p><ul><li><strong>Deleted Files:</strong> When a file is created, its entry with attributes is stored in the $I30. However, when the files are deleted, their metadata might still be preset in the $I30, which could help track deleting files until they are overwritten.</li><li><strong>Renamed / Moved File:</strong> Actions like renaming or moving files are also reflected in the $I30, which helps track file movement.</li><li><strong>Uncover Hidden Files:</strong> File Attributes can help identify the hidden files.</li></ul><h3>Analyzing $I30</h3><p>Let’s export this $I30 attribute file from the /[root]/metasploit directory using FTK Imager for the analysis. Once the $I30 index file is exported, we will use the following command to extract the information contained within the $I30 file, as shown below:</p><p>MFTECmd.exe -f ..\Evidence\$I30 --csv ..\Evidence\ --csvf i30.csv</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/841/1*lgNQQpa4pGeVKPCZ3AwXUg.png" /></figure><p>This command will parse and extract information from the $I30 attribute file.</p><p>Open the output file in the Timeline Explorer, as shown below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dgpOFzigx_BwAr7rqKuLPA.png" /></figure><h3>File Attributes in the $I30</h3><p>Some of the key information/metadata we can find within the $I30 about files and directories are listed below:</p><ul><li>Filename</li><li>From Slack</li><li>File Size</li><li>Parent Directory Information</li><li>Timestamps (MACB)</li><li>File Attributes</li><li>Read Only</li><li>Hidden</li><li>Directory / File</li><li>Archive / Compressed / Encrypted</li><li>System</li></ul><p>The output will show details about the files or directories present or present within the metasploit directory. We can apply a filter on the “From Slack” column to list down the files or directories that used to be present in this folder and are no longer present, hence being referred to in Slack Space.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*PKQqFGCKP5LzEMpqwkU3zw.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=17252a39ee5d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MBR and GPT Analysis]]></title>
            <link>https://medium.com/@ArkhamHunter/mbr-and-gpt-analysis-2c5ed0585ad0?source=rss-f52da4524c31------2</link>
            <guid isPermaLink="false">https://medium.com/p/2c5ed0585ad0</guid>
            <dc:creator><![CDATA[Abdullah Mohammed]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 13:23:17 GMT</pubDate>
            <atom:updated>2026-11-14T13:23:17.705Z</atom:updated>
            <content:encoded><![CDATA[<p>Imagine your disk as a large building storing all the data. This data is stored in binary format, as 1s and 0s, for the computers to understand. The challenge here is that without properly organizing this data, it would be a mess in the whole disk. To solve this problem, the disk is divided into multiple partitions, just like rooms in the building, and each partition contains specific data. For example, operating system files can be stored in one partition, personal files can be stored in another, etc. In the Windows OS, these partitions are represented by drive letters such as <strong>C</strong>, <strong>D</strong>, <strong>E</strong>, etc. Other operating systems may use different ways to refer to these partitions.</p><p>To solve this problem we need a map that helps us and to be guided through these partitions where they start and where they end and what they contain</p><ul><li>MBR → Master Boot Record</li><li>GPT → GUID Partition Table</li></ul><p>These two are different partitions schemas but have the same purpose but different arch and operation , choosing between these tow depends on multiple factors … including the disk size, hardware compatibility, and much more</p><p>The MBR/GPT is located in the very first sector of the disk and contains information about the structure and partitions of the disk</p><p><strong>It also plays a key role during the boot process of a system</strong></p><p>Due to this, the MBR/GPT has become an attractive target for <em>attackers</em> to manipulate the boot process by embedding their malware, often known as Bootkits, or tampering with them to make the system un-bootable</p><p>Let’s Take from the very beginning where we will talk about the Boot Process</p><p>this process wakes up the whole system</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/573/1*kobYfOFOma3TrUGJArB72g.png" /></figure><p>It starts by initializing the system’s hardware components, loading the operating system into memory, and finally allowing the user to interact with the system</p><p>Let’s take each one of them and discuss it .</p><h3>Power-On the System</h3><p>he first step of the boot process starts with pushing the power button, which sends electrical signals to the motherboard and initializes all the components</p><p>The CPU is the first component to get the electrical signals and needs some instructions to move further</p><p>we know that CPU executes instruction … from where does it fetches the instructions ?</p><p>it fetches the instructions from a chipset deployed on the motherboard and its known by “ BIOS/UEFI” and it contains instructions on how to get the boot process going.</p><p><strong>BIOS</strong> (Basic Input/Output System) and <strong>UEFI</strong> (Unified Extensible Firmware Interface) are responsible for verifying whether all the hardware components work properly.</p><p>A system can use either one of them , there are differences between them that lies in their capabilities and features</p><ul><li><strong>BIOS</strong> has been used for decades and is still used in some hardware. It runs in the basic 16-bit mode and supports only up to 2 terabytes of disks. The most important thing to note is that BIOS supports the MBR partitioning scheme</li><li><strong>UEFI</strong> came as a replacement for BIOS, offering 32-bit and 64-bit modes with up to 9 zettabytes of disks. UEFI offers a secure boot feature to ensure integrity during the system boot process. It also offers redundancy, allowing us to recover from the backup even if the boot code is corrupted. UEFI uses a GPT partitioning scheme, unlike the MBR partitioning scheme used by BIOS</li></ul><blockquote><em>To check if your system use BIOS or UEFI just press win+r and type </em><em>msinfo32 and go to BIOS mode field</em></blockquote><h3>Power-On-Self-Test (POST)</h3><p>Now the system is powered ON and the CPU already executed several instructions from the Firmware (BIOS/UEFI)</p><p>The BIOS/UEFI then starta a power-on-self-test to ensure theat all system’s hardware components are alright and ready , in case if there a component that it isnt working then the BIOS/UEFI will raise an error message on the screen for ex : “Keyboard Not Found “</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*RHAwsQ40xFbl2eANq56nCQ.png" /></figure><h3>Locate the Bootable Device</h3><p>After we pass the test the BIOS/UEFI searches for the bootable device ( the device that has the OS installed ) like SSD’S , HDD’S , USB</p><p>once the bootable device is located the BIOS/UEFI starts reading this device and here comes the role of MBR/GPT they will take control of the boot Process from here</p><blockquote><em>The method of checking your disk’s partitioning scheme will differ for different operating systems. For the Windows OS, you can type </em><em>Get-Disk in your PowerShell terminal, and it will show the partitioning schemes of the disks on your system.</em></blockquote><p>We have Been mentioning MBR and GPT , but what they consist of or what are those two partition schemas actually ?</p><p>let us talk first about the MBR</p><p>The bootable device which we discussed previously would be in the form of a disk , A disk is divided into multiple sectors and each with a standard size of 512 bytes</p><p>The first sector of this disk would contain the MBR</p><p>(A disk’s MBR can be viewed by taking the system’s disk image and opening it in a hexadecimal editor like HxD)</p><p>as we said The Master Boot Record (MBR) takes up 512 bytes of space at the very first sector of the disk</p><p>we can easily analyze the MBR code by starting from the first line , but how do we know where this MBR code ends ? the answer is straightforward</p><p>Every two digits coupled in hexadecimal represents 1 byte, and once the first 512 bytes of the disk completes, the MBR has been ended</p><p>Another way to spot the end is by looking at the MBR signature</p><p>the MBR signature is represented by <strong>55 AA</strong></p><p>Before we start analyzing the bytes, it’s important to understand that together, every two hexadecimal digits present a single byte. These 512 bytes of the MBR are further divided into three portions. The following screenshot highlights each of the three portions of the MBR with different colors</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*R2T2DbvD8irSVyO3FNHwfg.png" /></figure><p>The structure of the MBR can be seen below :</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xo7UpIe2-BiPnMS9LsorgQ.png" /></figure><p>Lets dissect each of the three portions of the MBR</p><h3>Bootloader Code (Bytes 0–445)</h3><p>the bootloader contains “ Initial Bootloader “ , the initial bootloader’s purpose is to find and locate the bootable partition from the partition table present in the MBR</p><h3>Partitions Table (Bytes 446–509)</h3><p>this table contains the details of all partitions present on the disk</p><p>One of the partitions in the disk contains all the operating system files necessary for booting, known as a <strong>bootable partition</strong></p><p>The initial bootloader that was started from the bootloader code of the MBR finds the bootable partition from this partition table</p><p>then the partition table loads the second bootloader and it loads the OS kernel in the memory</p><p>An MBR disk has a total of 4 partitions, and each partition is represented by 16 bytes in the partition table. In the screenshot below, four partitions (each with 16 bytes) are highlighted with different colors</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dZ3pRFAbh1ZHQXShcTvBPw.png" /></figure><p>The screenshot below shows the first partition from the partition table of the MBR. All the bytes have some specific meaning. Some bytes individually represent a field, while others are grouped together to form a field. We have highlighted all these different byte groups with different colors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/602/1*JfL1XpHHnS0vQdMfaDQxGQ.png" /></figure><p>Each field tells you something about the partition. Below is an explanation of all these fields :</p><ol><li><strong>Boot Indicator:</strong> This byte tells you whether the partition is bootable or not. A bootable partition contains files necessary for the operating system to boot. This boot indicator can only have one of the two values: 80 or 00. If it&#39;s 80, it means that the partition is bootable; else, if it&#39;s 00, it means that the partition is not bootable. In Windows based systems, <strong>C:</strong> is the partition that is typically bootable. If visualized through the partition table, this partition would have the boot indicator set to 80.</li></ol><p><strong>Starting CHS Address &amp; Ending CHS Address Those fields are not that important as we now have the simplified logical address of the partition in the form of the Starting LBA Address discussed ahead.</strong></p><ol><li><strong>Partition Type:</strong> Every partition uses a filesystem such as NTFS, FAT32, etc. This byte indicates the filesystem of the partition. The partition we are taking as a reference has this byte as 07, which means it is an NTFS partition. Every filesystem has its own unique byte. You can learn about the bytes used for other filesystems from <a href="https://www.writeblocked.org/resources/MBR_GPT_cheatsheet.pdf">here</a>.</li></ol><p><strong>3.Starting LBA Address:</strong> Logical Block Addressing (LBA) is the logical address that indicates the start of the partition. We saw that the Starting CHS Address also gives us the starting address of the partition, but because CHS gives you the physical address of the partition, it becomes difficult for us to locate it. However, the Starting LBA Address gives you the logical address of the partition rather than the physical address. You can use it to easily find the start of the partition on the disk in a hexadecimal editor. By locating the partition on the disk, you can also carve data from a disk’s hidden or deleted partitions. Based on the partition we are analyzing as a reference with 00 08 00 00 LBA</p><p><strong>4.Number of Sectors:</strong> These last 4 bytes of a partition tell you the number of sectors in the partition. We will calculate the sector’s size by using this field ahead.</p><h4>Locating the Partition</h4><p>The starting LBA address bytes of the partition we are using as a reference in this task is 00 08 00 00</p><p>these bytes are stored in the little-endian format in which the least significant byte is first and the most significant byte is last</p><p>we have to reverse it first 00 08 00 00 -&gt; 00 00 08 00</p><p>the next step is to convert these hexa bytes to decimal format ( you can use an online tool for this or view the decimal converted values of these bytes within the Hxd hex editor tool select the bytes for which you need the decimal value, and it will be displayed in the Int32 option of the Data Inspector pane on the right side)</p><p>now after converting it to decimal we have a decimal number → 2048</p><p>we have to multiply it with the size of the sector which is 512 bytes</p><p>2048 * 512 = 1,048,576</p><p>the last step is to go search with the byte offset 1000000 it will take you to the start of the partition on the disk</p><p>This can help you in carrying out detailed forensics of a specific partition. You can also recover any hidden or deleted data from here that has not yet been overwritten</p><blockquote><strong><em>Note:</em></strong><em> remember this method of reversing the LBA address (little endian), converting it to decimal, multiplying it with the sector size (512 bytes), and finally searching it into the hex editor to jump to the exact location. You will encounter some other Logical Block Addresses (LBA) in the GPT structure</em></blockquote><h4>Calculating the Size of Partition</h4><p>the last four partition bytes represent the number of sectors field .</p><p>As per our partition, these bytes are 00 B0 23 03 ( little-endian format)</p><p>we will reverse it so it will be → 03 23 B0 00</p><p>the next step is to convert it to decimal → 52,670,464</p><p>then multiply it with the size of a single sector which is 512</p><p>512 * 52,670,464 = 26,967,277,568 bytes which after rounding equals 26 GB</p><h3>MBR Signature (Bytes 510–511)</h3><p>the last part of the MBR is the MBR signature , its just two bytes</p><p>But if messed up , it can cause a huge trouble → the system cannot boot</p><p>these two bytes are 55 AA and are also known as Magic Number and they indicate that the MBR has ended</p><p>These bytes often get corrupted due to bad sectors on the disk and sometimes can be intentionally changed by malware to interrupt the normal boot process</p><p>to sum up the whole the boot process of the MBR :</p><ol><li>The initial bootloader starts from the bootloader code (the first component of the MBR).</li><li>It then finds the bootable partition from the partition table (the second component of the MBR).</li><li>Then it loads the second bootloader from this bootable partition.</li><li>Finally, with this, the OS kernel gets loaded, and after this process, the drivers, services, and filesystems are loaded into the memory, and the user is given control of the OS interface. This is a system’s whole boot process using an MBR partitioned disk</li></ol><h3>Threats targeting MBR</h3><p>MBR is not that common nowadays and is replaced by GPT . this happened when the UEFI firmware which is more efficient and redundant and compatible with GPT , replaced the BIOS which is compatible with MBR , however some disks still use the MBR it is important to know the threat surface of the MBR</p><p>As the MBR plays a crucial role in booting the operating system → it becomes an attractive target for attackers</p><p>The total 512 bytes of the MBR hold extreme importance for the whole system. Anybody who can play with these 512 bytes can hijack the whole system</p><p>Several types of malware are designed to target the MBR</p><ul><li>BootKits :</li></ul><p>Bootkits are a very dangerous type of malware. Since the MBR executes before the OS starts, the bootkits craft themselves into the MBR to bypass the OS’s protection mechanisms. Even complete malware removal or reinstallation of the OS cannot remove the bootkits as they reside in the MBR</p><ul><li>Ransomeware :</li></ul><p>MBR is the first thing executed on the disk during the boot process, and the boot process heavily relies on it. Instead of encrypting the individual files of the system, attackers encrypt and modify the MBR to interrupt the boot process and display their ransom notes on the screen. Petya ransomware, detected in 2016, encrypts the MBR rather than all the disk files. Another example is the Bad Rabbit ransomware, which overwrites the original MBR with its malicious bootloader code and displays a ransom note on the screen upon reboot of the system</p><ul><li><strong>Wiper Malware :</strong></li></ul><p>This type of malware intends to corrupt the MBR to make the system unbootable. Any change in the original MBR of a disk can make the system completely unbootable. The Shamoon malware is one such malware that overwrites the MBR with random characters to make the system unbootable. In the next task, we will examine a similar case in which malware corrupted the MBR of a disk to make the system unbootable</p><p>Now Lets Talk about the GPT</p><p>Previously, we studied how the boot process propagates with the MBR , now we will explore how the process will look like when GPT partitioning scheme is used in the disk instead of the MBR</p><p>UEFI firmware with the GPT partitioning scheme replaced BIOS firmware with the MBR partitioning scheme</p><p>This replacement occurred due to some limitations in the BIOS and the MBR. The GPT supports up to 9 zettabytes of hard disks, unlike the MBR, which supports a maximum of 2 terabytes. The GPT also supports up to 128 partitions, unlike the MBR, which only supports 4. There are some other differences between both of the partitioning schemes.</p><p>We previously saw that the MBR only occupies the disk’s first sector (512 bytes). However, this is not the case with the GPT. The GPT has five components spread across multiple sectors of the disk. Let’s start dissecting each component of the GPT.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*-PbLxtCjLprpx7oG-ijmXw.png" /></figure><h3>Protective MBR</h3><p>What if the disk it GPT Partitioned and the system use BIOS firmware even though the disk is designed to work with the GPT ?</p><p>to solve this problem the GPT has Protective MBR , it is located in the first sector of the disk and its purpose is to signal the BIOS System and applications that the disk is using GPT , so please dont mess up with it thinking it is the MBR</p><p>he Protective MBR also has three components, just like the general MBR, but with a few differences</p><p>Below is a screenshot of the Protective MBR found in the GPT</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7sqTepuDxGGkvNjssCi8rQ.png" /></figure><p>The details of these components are given below :</p><p><strong>1.Bootloader Code:</strong> This bootloader code is not the same as it is in the general MBR. This bootloader code does not perform any function during the boot process. It is just there to look like it’s the same standard MBR bootloader. This would be all 00s in most scenarios; however, sometimes, this can contain some placeholder code for legacy compatibility.</p><p><strong>2.Partition Table:</strong> This partition table contains only one partition (the first 16 bytes), and this partition has one job; to redirect the system to the EFI Partition (which we will discuss later). The screenshot of the protective MBR above shows that it only has one partition in the table, and the other partitions are labeled with 0s. In this single partition, there is only one important thing: the 4th byte. This byte is set to EE, indicating that this is a GPT-formatted disk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WdjXxXjukEMxktPI50O1zg.png" /></figure><p>3. MBR Signature: The MBR signature is the same as in the standard MBR. It is set to 55 AA and marks the end of the Protective MBR.</p><h3>Primary GPT Header</h3><p>The GPT header starts right from the next byte (the start of sector 1) after the Protective MBR ends at 55 AA (the end of sector 0)</p><p>All the bytes in the GPT header have a specific meaning. Below is a screenshot of the GPT header with these bytes highlighted with different colors. It has the whole 512 sector space but occupies the first 92 bytes of space, and after these 92 bytes, there would be all 00 bytes for padding purposes to complete the sector&#39;s total bytes. So, for the Primary GPT header, you only need to focus on the first 92 bytes, which can give you some meaningful information.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KdyD_O1Lx9kcu4KPx6E03w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/666/1*htocNRNwg3Mucmcm9QQSCg.png" /></figure><p>Each of these 14 fields has a specific purpose and meaning , Below is an explanation of all these fields , Before starting with the explanation of these fields, it is important to note that some of the fields represent Logical Block Addresses (LBA). To calculate the exact address through these LBAs, you must follow the same steps we performed while locating the partition in the MBR</p><p>These steps included reversing the bytes stored in a little-endian format, then converting them to decimals and searching them on the HxD tool to jump to the exact LBA location</p><ol><li><strong>Signature:</strong> This field has a value 45 46 49 20 50 41 52 54 which recognizes it as a GPT header. This value is always at the start of the GPT header.</li><li><strong>Revision:</strong> The revision number is of 4 bytes and it represents the version of the GPT. Most of the times it would be 00 00 01 00 which means the GPT version is 1.0.</li><li><strong>Header Size:</strong> This field represents the size of the GPT header. It is typically 5C 00 00 00 in hex and if you convert it to decimal (after reversing the order of bytes as they are in little-endian), it is 92 bytes which is the length of the GPT header.</li><li><strong>CRC32 of Header:</strong> This is the CRC32 checksum of the GPT header, which if changed, would indicate that either the GPT header is tampered or corrupted.</li><li><strong>Reserved:</strong> These are reserved bytes. The purpose of having them is to utilize them for any future changes in the GPT header.</li><li><strong>Current LBA:</strong> The Current Logical Block Address (LBA) indicates the location of the GPT header. We know that its location is in sector 1, and we can verify this by converting the 8 Current LBA bytes 01 00 00 00 00 00 00 00 into decimal.</li><li><strong>Backup LBA:</strong> In the GPT partitioning scheme, we have a backup of the GPT header as well, which we will be studying later on in this task. This field indicates the LBA of the backup GPT header.</li><li><strong>First Usable LBA:</strong> This LBA address indicates the first address from which the partition can start on the disk.</li><li><strong>Last Usable LBA:</strong> This LBA address indicates the last address to which the partitions on the disk can be written. Any partitions cannot occupy the disk space after the last usable LBA.</li><li><strong>Disk GUID:</strong> This field is of 16 bytes and it presents a Globally Unique Identifier of the disk. The purpose of this GUID is to distinguish the disk from any other disks present in the system. In the current GPT header that we are analyzing, these bytes are 1D F1 B0 D6 43 BE 37 4E B1 E6 38 66 EC B1 73 89. We can convert them to the standard GUID format of the disk by just reformatting them as 1DF1B0D6-43BE-374E-B1E6-3866ECB17389.</li><li><strong>Partition Entry Array LBA:</strong> This LBA address indicates the start of the Partition Entry Array which we are going to discuss ahead as the 3rd component of the GPT.</li><li><strong>Number of Partition Entries:</strong> This field indicates the number of partitions that are on the disk. The GPT supports 128 partitions, unlike the MBR, which supports 4 partitions only. The value of this field is 80 00 00 00 which if converted to decimal will be 128.</li><li><strong>Size of Each Partition Entry:</strong> This field indicates the size occupied by each partition entry array. In this example, it set to 80 00 00 00 which is 128 in decimal. It is important to note that this is not the size of the partition itself. This is just the size of partition entry array that we would be discussing next.</li><li><strong>CRC32 of Partition Array:</strong> This is the CRC32 checksum of the whole partition entry array, which if changed, would indicate that either the partition entry array is tampered or corrupted.</li></ol><h3>Partition Entry Array</h3><p>We saw that sector zero was occupied by the Protective MBR, and the GPT header occupied sector 1. Now, from sector 2</p><p>the Partition Entry Array starts, just like the partition table present in the MBR, with a few differences. There are a total of 128 partitions on a GPT disk</p><p>and this partition entry array contains information about all these partitions</p><p>Below is the screenshot of the Partition Entry Array of a GPT disk. Each partition entry is represented by 128 bytes. You can only see the 6 partition entries out of the total 128 partition entries of the GPT. This is because there are only six working partitions in this disk. These six partitions would be present in blocks (128 bytes each) in this partition entry array, and after these working partitions, all the remaining 122 partition entries would be marked with 00.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hrjzKB5pFPaQn2wvJ34fKg.png" /></figure><p>Let’s use the first partition of the above screenshot as a reference to analyze the meaning of all the hexadecimal digits present in it</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/594/1*_E5pfINthAsFe6d7FcMYGg.png" /></figure><p>Each of these six fields has a specific purpose. Below is an explanation of all these fields.</p><ol><li><strong>Partition Type GUID:</strong> This is the GUID of the partition type. This GUID will indicate the partition type, i.e., EFI System Partition, Basic Data Partition, etc. The 16 bytes of our partition entry are stored in mixed endian (little-endian and big-endian) format. This means we would have to reverse specific bytes and keep the other ones the same. In this case, we would do the following:</li><li>Reverse the first 4 bytes from 28 73 2A C1 to C1 2A 73 28, as they are in little-endian format.</li><li>Reverse the next 2 bytes from 1F F8 to F8 1F, as they are in little-endian format.</li><li>Reverse the next 2 bytes from D2 11 to 11 D2 as they are in little-endian format.</li><li>Keep the next 2 bytes BA 4B as it is, as they are in big-endian format.</li><li>Keep the last 6 bytes 00 A0 C9 3E C9 3B as it is, as they are in big-endian format.</li></ol><p>The resulting GUID from steps 1–5 will be C12A7328-F81F-11D2-BA4B-00A0C93EC93B. Whatever GUID you get from here, you can search it on the internet to get the partition type associated with that GUID. The GUID we calculated is used for EFI System Partition (ESP).</p><p><strong>EFI System Partition (ESP)</strong> is stored as a partition on every GPT disk and it is a very critical component in the boot process. In the MBR, we saw that the bootloaders are located in the bootcode and the bootable partition. In contrast, in the GPT, the bootloader is comprised of multiple files with <strong>.efi</strong> extension, and they all are stored in this EFI System Partition (ESP). Once the bootloader is found from the EFI System Partition, it loads the <strong>Operating System’s kernel</strong> from the bootable partition of the disk, and after that, the drivers, services, and filesystems are loaded into the memory. Then the user is given the control of OS interface.</p><ol><li><strong>Unique Partition GUID:</strong> Unique Partition GUID is used to distinguish partitions on a disk. It is a unique GUID that is given to all the partitions on the disk. To convert these hexadecimal bytes into the standard GUID format, you can follow the same steps as we did for the first field (Partition Type GUID), as this is also stored in the mixed endian format.</li><li><strong>Starting LBA:</strong> The starting LBA address indicates the area from where this partition starts on the disk.</li><li><strong>Ending LBA:</strong> The ending LBA address indicates the area at which this partition is ending on the disk.</li><li><strong>Attributes:</strong> This field contains some flags that indicates some features of the partition, for example, if it is bootable, hidden, or normal.</li><li><strong>Partition Name:</strong> This is the last field of the partition entry, and its size is 72 bytes. It represents the name of the partition in string format and is UTF-16 encoded. If you decode these bytes using any online hex-to-string decoder, you will get the partition name of this partition.</li></ol><h3>Backup GPT Header</h3><p>one of the the biggest reasons the GPT replaced the MBR is its redundancy . in case of the MBR , if your MBR gets corrupted or changed due to hardware issues or malicious attacks ,you have a minimal chance of recovery .</p><p>On the other hand, the GPT has backups of its components</p><p>Backup GPT Header is located at the last sector of the disk, and it can be used for recovery in case the Primary GPT Header is corrupted. It contains the same information as the Primary GPT Header.</p><h3>Backup Partition Entry Array</h3><p>there is also a backup copy of the partition entry array</p><p>It is stored at the end of the disk just before the backup GPT header.</p><p>It contains the same information as the partition entry array. It acts as a fail-safe mechanism in case the original partition entry array is damaged</p><h3>Threats Targeting GPT</h3><p>GPT offers many advantages over the MBR in terms of efficiency and redundancy</p><p>However, it is also prone to some attacks</p><ul><li>Bootkits :</li></ul><p>GPT has its bootloader code inside the EFI System Partition (ESP) and is managed by UEFI firmware. This bootloader code is in the ESP’s multiple <strong>.efi</strong> extension files. Attackers target these <strong>.efi</strong> files and replace them with their malicious bootkits. This allows them to run before the OS boots and bypass the OS-level protection mechanisms. The Secure Boot feature of UEFI can prevent this from happening as it verifies the integrity of the bootloader files by their digital signatures</p><ul><li>Ransomware :</li></ul><p>Ransomware can also target the GPT to disrupt the booting process. This attack can be difficult compared to the MBR as the redundancy provided by the GPT in the form of the <strong>backup</strong> GPT header and partition entry array prevents this. However, some advanced malware may encrypt the ESP, which can result in the disruption of the boot process.</p><ul><li>Wiper malware :</li></ul><p>GPT provides a backup header and partition entry array, so the system can recover through these backups even if the malware destroys the primary ones. However, some advanced malware uses techniques to encrypt both the primary and the backup GPT header and partition entry array to disrupt the boot process completely. Some malware may target the EFI system partition (ESP) directly, making the recovery nearly impossible.</p><p>When we use BIOS firmware, the bootloader responsible for loading the operating system is executed inside the MBR. However, when we use UEFI firmware, which supports a GPT partitioning scheme, the bootloader is not executed from the GPT. Instead, it is located in the EFI system partition (ESP) in the form of <strong>.efi</strong> extension files. These files include the <strong>bootmgr.efi</strong>, <strong>bootx64.efi</strong>, and others, depending on the OS.</p><p>Earlier, we studied the bootloader code in the MBR, which can be tampered with, and a bootkit can reside there. Conversely, using UEFI firmware with the GPT, bootkits can reside in the .efi files in the ESP partition. However, this can only be successful if the secure boot feature of UEFI is disabled. Once enabled, the secure boot feature will verify the digital signatures of the bootloaders (.efi files) in the ESP and will not allow any tampered file to execute.</p><p>Hope you enjoyed our first footsteps in Forensics folks ..</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*PKQqFGCKP5LzEMpqwkU3zw.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2c5ed0585ad0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Carnage — THM]]></title>
            <link>https://medium.com/@ArkhamHunter/carnage-thm-fdbe83291b12?source=rss-f52da4524c31------2</link>
            <guid isPermaLink="false">https://medium.com/p/fdbe83291b12</guid>
            <dc:creator><![CDATA[Abdullah Mohammed]]></dc:creator>
            <pubDate>Sun, 04 Oct 2026 12:24:42 GMT</pubDate>
            <atom:updated>2026-10-04T12:24:42.964Z</atom:updated>
            <content:encoded><![CDATA[<h3>Scenario</h3><blockquote><em>Eric Fischer from the Purchasing Department at Bartell Ltd has received an email from a known contact with a Word document attachment. Upon opening the document, he accidentally clicked on “Enable Content.” The SOC Department immediately received an alert from the endpoint agent that Eric’s workstation was making suspicious connections outbound. The pcap was retrieved from the network sensor and handed to you for analysis.</em></blockquote><blockquote><strong><em>Task</em></strong><em>: Investigate the packet capture and uncover the malicious activities.</em></blockquote><p>⚠️ → DO NOT directly interact with any domains and IP addresses in this challenge</p><p>1- What was the date and time for the first HTTP connection to the malicious IP?</p><p>(<strong>answer format</strong>: yyyy-mm-dd hh:mm:ss)</p><p>apply the https filter and take take the timestamp from the 1st packet</p><p>2- What is the name of the zip file that was downloaded?</p><p>apply the http filter and the packet will appear for the downloaded document</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MZ3n0k5vfyclYsIBKf-24w.png" /></figure><p>3- What was the domain hosting the malicious zip file?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RCRa2KZJucPinWRS_GSbug.png" /></figure><p>4- Without downloading the file, what is the name of the file in the zip file?</p><p>follow the tcp stream for the packet and inspect deeply the stream and u will find the file</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vv-uDViw111oYTHszfZm0A.png" /></figure><p>5-What is the name of the webserver of the malicious IP from which the zip file was downloaded?</p><p>like the privious TASK follow the tcp stream and go to the server field</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e8CFpOOhVUU3vaDu-aEZvg.png" /></figure><p>6- What is the version of the webserver from the previous question?<br> from the same page head to x-powered by Field</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tNgyz9IK0BXEYL6vPCehsA.png" /></figure><p>7-Malicious files were downloaded to the victim host from multiple domains. What were the three ****domains involved with this activity?<br> apply the dns filter and u will find 3 suspicious domains</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rAHAtWwakv41u-EAj_pc-A.png" /></figure><p>8-Which certificate authority issued the SSL certificate to the first domain from the previous question?</p><p>follow the tcp packet of the <a href="http://finejewels.com.au/">finejewels.com.au</a> and deeply inspect the stream till you find the Secure Certificate Authority field</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w_UhrToJoGt5NqT7YW2dgA.png" /></figure><p>9 — What are the two IP addresses of the Cobalt Strike servers? Use VirusTotal (the Community tab) to confirm if IPs are identified as Cobalt Strike C2 servers. (answer format: enter the IP addresses in sequential order)</p><p>click on the tcp tab and then click on the packets tab to get the highest packets sent<br> here u could use your research but here are 2 ip’s if we search in VT we will see that its associated with Cobalt Strike</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KKkDdXvuxVVnBSq7oRFYeg.png" /></figure><p>10- What is the Host header for the first Cobalt Strike IP address from the previous question?</p><p>now we know the malicious ip address we could apply it as a filter and the follow the tcp stream and the host header will appear</p><p>11- What is the domain name for the first IP address of the Cobalt Strike server? You may use VirusTotal to confirm if it’s the Cobalt Strike server (check the Community tab).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/729/1*Lvjn-UA6sQfePd5nJJCbOQ.png" /></figure><p>12- What is the domain name of the second Cobalt Strike server IP? You may use VirusTotal to confirm if it’s the Cobalt Strike server (check the Community tab).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/690/1*oqPJ-kNsaeoPQt4NLevamQ.png" /></figure><p>13- What is the domain name of the post-infection traffic?</p><p>apply the HTTP.request.method == POST</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ugOEFwQUbMrGnIeUgU0kSA.png" /></figure><p>14 — What are the first eleven characters that the victim host sends out to the malicious domain involved in the post-infection traffic?</p><p>from the same page</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kiyz0XTts6fCCcisWq_uLg.png" /></figure><p>15 — What was the length for the first packet sent out to the C2 server?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lQK2WBIuyaXV9rpDEAZyZA.png" /></figure><p>16 — What was the Server header for the malicious domain from the previous question?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qjqC7d4YzpdBIdWTevAJFQ.png" /></figure><p>17- The malware used an API to check for the IP address of the victim’s machine. What was the date and time when the ****DNS ****query for the IP check domain occurred? (<strong>answer format</strong>: yyyy-mm-dd hh:mm:ss UTC)<br> since it focuses on api we will search for a domain that contain and uses an api</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qjqC7d4YzpdBIdWTevAJFQ.png" /></figure><p>18 — What was the domain in the DNS query from the previous question?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b1tu8Yku8xfEOF4t9yYK7Q.png" /></figure><p>19 — Looks like there was some malicious spam (malspam) activity going on. What was the first MAIL FROM address observed in the traffic?</p><p>since its and email then we will apply the smtp filter and look for the first MAIL FROM address</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DjgBHam-u7ZkSKghdBPuzQ.png" /></figure><p>20 — How many packets were observed for the SMTP ****traffic?</p><p>from the same Page</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DjgBHam-u7ZkSKghdBPuzQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*PKQqFGCKP5LzEMpqwkU3zw.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fdbe83291b12" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Agent Discovery — THM]]></title>
            <link>https://medium.com/@ArkhamHunter/agent-discovery-thm-285d37b915a8?source=rss-f52da4524c31------2</link>
            <guid isPermaLink="false">https://medium.com/p/285d37b915a8</guid>
            <dc:creator><![CDATA[Abdullah Mohammed]]></dc:creator>
            <pubDate>Mon, 14 Sep 2026 14:00:23 GMT</pubDate>
            <atom:updated>2026-09-14T14:00:23.368Z</atom:updated>
            <content:encoded><![CDATA[<p>Task 1 — Introduction</p><p>Sets the scenario and stakes: the proposed system, whatever it ends up being, must never take containment actions itself — no closing alerts, blocking IPs, disabling accounts, or touching security controls. The engineer stays the final decision-maker regardless of what gets built. No graded question beyond confirming readiness — the quality of your reasoning matters more than which of the five outcomes you land on.</p><p>Task 2 — Workflows, Assistants, and Agents</p><p>Establishes the core taxonomy the entire room hinges on — distinguished by who controls the next step, and whether an LLM is involved at all:</p><p>Traditional workflow — fixed code, rules, or scripts, no LLM. The developer defines the sequence up front; the same input always produces the same action.</p><p>AI workflow — an LLM operating inside a sequence that code still controls. The model might summarise or interpret, but the developer decides the order of steps and the branch/stop conditions.</p><p>AI assistant — the human asks, the assistant answers. The engineer decides what to ask and what to do with the answer; the assistant never independently drives the investigation.</p><p>AI agent — given a goal, the model dynamically decides how to proceed: which approved tool to reach for, when to consult documentation, whether another search is needed, how to adjust its plan based on what it finds.</p><p>Anthropic’s own engineering guidance is referenced here directly: start with the simplest composable approach and add complexity only when it produces a real improvement — workflows buy consistency on well-defined tasks, while agents make sense specifically when the required steps can’t be predicted ahead of time (at the cost of more latency, cost, and compounding error risk).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/971/1*T5diZ5qWRmeXCKImjUd4LA.png" /></figure><p>Task 3 — Why Agent Discovery Matters</p><p>The room’s central argument: an AI agent is not the goal — improving the actual workflow safely is. “NorthStar Fashion needs an AI agent” is a proposed solution, not a description of the problem; starting from a preferred technology risks producing something technically impressive that still misses the real bottleneck.</p><p>Key distinction: repetitive doesn’t mean agentic. Deterministic tasks (extracting an IP, sorting timestamps by rule) belong in a traditional workflow regardless of how repetitive they are. Agentic behaviour earns its place specifically when the next step genuinely can’t be predicted in advance — e.g. the model needs to pick a different investigation tool because the first one came back with weak evidence.</p><p>The room also lists concrete failure modes discovery is meant to prevent: automating the wrong bottleneck, reaching for AI where plain rules would do, assuming underlying documentation is clean and ready when it isn’t, and failing to define what “improvement” even means up front. Agents specifically demand extra scrutiny because flexibility (choosing tools, changing plans mid-investigation) raises real questions — what can it access, how does it handle an unreliable source, when must it stop and hand off to a human — that a fixed workflow never has to answer, since a plausible-sounding LLM output is not the same thing as a correct one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/957/1*qrNQXuCn8-rZfPI37ZhiDA.png" /></figure><p>Task 4 — Understanding the Business Problem</p><p>Turns NorthStar’s vague request (“build an AI agent to help investigate alerts”) into an actual, evidence-backed problem statement — deliberately before deciding on any of the five outcomes. The evidence gathered:</p><p>Alert volume meets limited capacity — ~50 alerts/day is manageable for a dedicated team, but NorthStar has exactly one engineer wearing every hat.</p><p>High severity can be a false positive — a SIEM flagged repeated scanning as high severity; it turned out to be an approved vulnerability scanner, only discoverable by checking prior documentation.</p><p>Low severity can hide real risk — the mirror-image failure: an alert scored low severity had a source IP that an outside reputation check showed was tied to malicious activity.</p><p>Related alerts change meaning together — an unusual-location login followed ten minutes later by an external mail-forwarding rule; each looks low/medium risk alone, together they read as account compromise.</p><p>Evidence is distributed — SIEM alert, IP reputation, related events, past incident notes, personal notes — the engineer has to manually pull all of it together.</p><p>This synthesises into a formal business problem statement: NorthStar’s single engineer must manually gather and interpret evidence across several disconnected sources while triaging ~50 alerts a day, SIEM severity alone is unreliable, and related events can flip an alert’s real meaning — all of which eats into time needed for the engineer’s other responsibilities. Notice what the statement deliberately does not do: it doesn’t prescribe a workflow, assistant, or agent as the fix.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/855/1*CtYpqsHSJbwLOaoodofCOw.png" /></figure><p>Task 5 — Mapping the Workflow</p><p>Where Task 4 established what’s wrong, this task maps how the work actually happens today — deliberately still not designing anything new yet. The exercise (sometimes called “workflow archaeology,” since some steps live only in the engineer’s habits, not any official doc) sets clear boundaries: starts when a SIEM alert is available, ends when the engineer records a conclusion and next-step priority. Containment actions are explicitly out of scope for the whole exercise.</p><p>The current-state workflow, in eight steps: (1) review the SIEM alert → (2) decide what evidence is actually needed, since not every alert carries the same fields → (3) check IP reputation if a source IP exists → (4) search prior incident docs/personal notes for matching context → (5) search for related alerts sharing the same user/IP/device/timeframe → (6) compare and correlate everything gathered (the most judgement-heavy step) → (7) decide: deprioritise, investigate further, or escalate — looping back to an earlier step if the evidence is thin or contradictory → (8) record the outcome and the reasoning behind it for future reference.</p><p>No graded question in this task beyond the conceptual walkthrough — it exists purely to produce the artefact (the workflow map) that Task 6’s canvas exercise runs on.</p><p>Task 6 — Agent Discovery Canvas (Practical)</p><p>An interactive static site (opened via “View Site”) where you enter the workflow map you just built, evaluate each task’s frequency/complexity/risk, apply something the room calls the Human Agency Scale, and compare all five solution paths side by side before landing on a final recommendation. Completing the full exercise across its stages reveals the flag.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/947/1*Orhv9NP8OQdjc4qJQ_PSmw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*PKQqFGCKP5LzEMpqwkU3zw.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=285d37b915a8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Securing the AI Supply Chain — THM]]></title>
            <link>https://medium.com/@ArkhamHunter/securing-the-ai-supply-chain-thm-399751a00d37?source=rss-f52da4524c31------2</link>
            <guid isPermaLink="false">https://medium.com/p/399751a00d37</guid>
            <dc:creator><![CDATA[Abdullah Mohammed]]></dc:creator>
            <pubDate>Sun, 13 Sep 2026 13:52:56 GMT</pubDate>
            <atom:updated>2026-09-13T13:52:56.199Z</atom:updated>
            <content:encoded><![CDATA[<p>Starts at the root cause from the previous room: pickle’s ability to execute arbitrary code on load. Two independent fixes:</p><p>SafeTensors — the serialisation format Hugging Face created specifically to replace pickle for ML models; it stores only tensor data, with no mechanism for embedding executable objects.</p><p>`weights_only=True` — a PyTorch loading parameter that, even when you’re still stuck with pickle-based `.pt`/`.pth` files, restricts deserialisation to tensor data only and blocks arbitrary object reconstruction — a compensating control when you can’t convert the format outright.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/930/1*IR5AcA1-T3EOSA2E_JgE5w.png" /></figure><p>Task 2 — The Model Acquisition Framework and Integrity Verification</p><p>Introduces a formal intake process for any new model entering the pipeline — the Model Acquisition Framework — whose first step, before anything else happens to a newly received model, is Quarantine: isolating it from production systems until it’s been verified.</p><p>Verification itself starts with checksums — comparing a model file’s hash against its expected/published value to catch tampering.</p><p>Practical: examine checksums across the lab VM’s model files; the file that fails to match its expected hash is `model_review_v2.pkl`.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*alq7AR-px8A3chuVrtRiKQ.png" /></figure><p>Task 3 — Telemetry and Runtime Behaviour</p><p>Even a model that passes a checksum check can still be malicious if the checksum itself was computed after tampering, or if the check was skipped. This task uses runtime telemetry to catch what static checks miss: loading the compromised model, its telemetry shows the object returned on “load completion” is an `int` — not a model object at all, a dead giveaway that the pickle payload replaced normal model loading with something else entirely (almost certainly the return value of whatever malicious code the payload executed).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/845/1*NrgkB0Lm9TfrYe-bjsm3dg.png" /></figure><p>Task 4 — Static Scanning Tools</p><p>Introduces two purpose-built scanners for exactly this class of threat:</p><p>Fickling — Trail of Bits’ static analysis tool for pickle files; it can inspect a pickle’s opcode stream (building on the `pickletools` skill from the previous room) and flag dangerous operations without executing them.</p><p>ModelScan — a broader model-scanning tool that assigns severity ratings to suspicious operations found inside model files; it flags an `os.system` call in a model file as CRITICAL severity, reflecting that such a call is a near-certain sign of an arbitrary-code-execution payload.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/819/1*mGfnOQCDOllmb11dhNVz5g.png" /></figure><p>Task 5 — Architecture Inspection</p><p>Returns to the telemetry terminal to inspect the compromised model’s actual architecture layer-by-layer. The model’s architecture contains 5 layers — walking through them is how you’d spot an injected layer that doesn’t belong in the model’s intended design (the kind of architecture-level attack that SafeTensors conversion alone can’t catch, as established in the previous room).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/778/1*VoRimg8qKk0CqcNBL9foCw.png" /></figure><p>Task 6 — Finding Malicious Layers</p><p>Applies architecture inspection to a second file — `image_classifier_v2.h5` — using `inspect_h5_model.py`. This surfaces a suspicious Lambda layer (the same layer type flagged as surviving SafeTensors conversion in the previous room) named `manipulate_output` — a name that telegraphs exactly what it’s doing: altering the model’s output after inference rather than contributing to legitimate classification.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/903/1*VaKpMwARGwJDCb_GD2mj4w.png" /></figure><p>Task 7 — Dependency Auditing and SBOMs</p><p>Moves from model files to the dependency layer (the `torchtriton`-style attack from earlier rooms). Three practices:</p><p>Version pinning — the recommended practice for specifying package versions in `requirements.txt`, so an unexpected transitive update (or a malicious package publish) can’t silently get pulled in.</p><p>`pip-audit` — scans installed Python dependencies against known vulnerability databases.</p><p>CycloneDX — the OWASP-maintained SBOM (Software Bill of Materials) format focused specifically on security use cases, letting you track exactly what’s in your dependency tree and cross-reference it against advisories.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/899/1*-sMo8wJOeA75IJk8NhcGYw.png" /></figure><p>Task 8 — Governing API-Consumed Models and Prompt Templates</p><p>Closes the loop on the API supply chain risks introduced in the previous room:</p><p>Behavioural baseline — what you need to establish in order to detect a silent model update behind a hosted API endpoint; without a documented baseline of expected behaviour, a swapped model underneath the same endpoint is invisible.</p><p>System prompts as a governed artefact — the room’s answer to “what supply chain artefact should be version-controlled and reviewed like code” is system prompts specifically, since (as covered previously) an untrusted prompt template can alter LLM behaviour across every application that reuses it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/932/1*Y1rUV2Rmp-L7ljLLAegeig.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*PKQqFGCKP5LzEMpqwkU3zw.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=399751a00d37" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Supply Chain Attack Vectors — THM]]></title>
            <link>https://medium.com/@ArkhamHunter/supply-chain-attack-vectors-thm-b9554382ccf7?source=rss-f52da4524c31------2</link>
            <guid isPermaLink="false">https://medium.com/p/b9554382ccf7</guid>
            <dc:creator><![CDATA[Abdullah Mohammed]]></dc:creator>
            <pubDate>Fri, 11 Sep 2026 11:31:12 GMT</pubDate>
            <atom:updated>2026-09-11T11:31:12.338Z</atom:updated>
            <content:encoded><![CDATA[<p>Task 1 — Pickle Serialisation and Code Execution</p><p>Covers the root cause of most model-file-based attacks: Python’s pickle format. When pickle deserialises a custom object, it calls the object’s `__reduce__` method to get reconstruction instructions — and an attacker can override `__reduce__` to return a callable and arguments of their choosing, which pickle then executes automatically on load. The classic payload abuses Python’s built-in `os` module (e.g. `os.system`) to run arbitrary shell commands the moment the “model” is loaded.</p><p>Also covers the limits of the fix from the previous room: converting a Keras model to SafeTensors removes the pickle payload, but leaves architecture-level attacks completely untouched — because SafeTensors only changes how weights are stored, not what the model’s architecture does. A concrete example: converting a `.h5` Keras model to SafeTensors does not remove a malicious Lambda layer, since that’s a functional part of the model graph, not a serialisation artefact</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*84uIBxRevV-DkqWQlndoUg.png" /></figure><p>Task 2 — Safe Analysis of a Malicious Model</p><p>Introduces `pickletools` — a Python module that can safely disassemble a pickle file into its opcodes without ever executing it, letting you inspect what a suspicious model file would actually do before loading it for real.</p><p>Practical: analyse a malicious model file from the attached target VM using `pickletools`, and identify the external domain the payload tries to contact, plus the specific opcode responsible for triggering execution. The `REDUCE` opcode is what actually calls the function named by a preceding `STACK_GLOBAL` opcode — that’s the exact mechanical step where deserialisation turns into code execution</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/890/1*NPzlK-nRhhhPI_jRrSfPrw.png" /></figure><p>Task 3 — Dependency Confusion</p><p>Covers the second attack layer from the previous room in more depth: dependency confusion — the term for an attack where a package published on a public registry overrides an internal, same-named package because of how dependency resolvers prioritise sources. If an internal package name isn’t also registered publicly, an attacker can claim that name on PyPI (or npm) and get it installed instead of the real internal one, purely by exploiting resolution order</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qKJJg7VDLeE-69bEKq7RhA.png" /></figure><p>Term for an attack where a public package overrides an internal package of the same name? Dependency confusion</p><p>Task 4 — Typosquatting</p><p>Covers the sibling technique to dependency confusion: typosquatting — publishing packages or models under names deliberately close to legitimate ones (a single character swapped, transposed, or added), betting that a developer will mistype or not look closely enough to notice.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/889/1*_8u-CWm5jhaNyCVEO6vHsA.png" /></figure><p>Task 5 — Repository Manipulation and the TryTrainMe Breach</p><p>Ties the investigation together. The attacker created a Hugging-Face-style organisation called `trustworthy-ai-lab` specifically to make a malicious model download look legitimate — a textbook example of repository manipulation, the third of the three attack vectors covered in this room’s reference table (alongside malicious model files and dependency confusion). The room then asks you to reason about defence-in-depth: even if TryTrainMe’s model loader had successfully blocked the pickle payload, dependency confusion was still sitting there as a second, independent path to code execution — meaning fixing one vector wouldn’t have stopped the breach</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N-3mmHBOZfQI8hwqd9npdQ.png" /></figure><p>Task 6 — API Supply Chain Risks</p><p>Shifts focus to organisations that consume AI via hosted APIs (OpenAI, Anthropic, Google, or aggregators) rather than running models locally. Two risks specific to that model:</p><p>Silent model updates — the risk that the model actually serving a given API endpoint changes without the consumer being told, meaning behaviour (and safety properties) you tested against can shift underneath you without warning.</p><p>Prompt templates as a supply chain artefact — if pulled from an untrusted repository, a shared prompt template can alter an LLM’s behaviour consistently across every application that reuses it, making it a high-leverage, easily overlooked attack surface</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/915/1*2OUlqmdVBJqSXZCJESlymA.png" /></figure><p>Practical: Interrogating TryAssist</p><p>A live-agent exercise against TryAssist, TryTrainMe’s internal AI assistant, probing it about its own governance practices as part of the investigation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1004/1*7mDkdEdaGLPN04O9bPmsOg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*PKQqFGCKP5LzEMpqwkU3zw.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b9554382ccf7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Understanding AI Supply Chains — THM]]></title>
            <link>https://medium.com/@ArkhamHunter/understanding-ai-supply-chains-thm-bce2a00fc809?source=rss-f52da4524c31------2</link>
            <guid isPermaLink="false">https://medium.com/p/bce2a00fc809</guid>
            <dc:creator><![CDATA[Abdullah Mohammed]]></dc:creator>
            <pubDate>Tue, 08 Sep 2026 11:09:58 GMT</pubDate>
            <atom:updated>2026-09-08T11:09:58.517Z</atom:updated>
            <content:encoded><![CDATA[<p>Task 1 — Introduction</p><p>Sets the scenario and lists what you’ll walk away with: understanding what an AI supply chain is, why it differs from a traditional software supply chain, and where it gets attacked. No graded question beyond confirming readiness.</p><p>Task 2 — What Is a Supply Chain?</p><p>Starts with the traditional software supply chain definition before layering AI-specific concerns on top. Covers the SolarWinds attack as the canonical software supply chain incident — the malicious code was injected during the build process, not into the final shipped source, which is why it went undetected for so long (it compromised the trusted step that turns source into a trusted binary).</p><p>Also introduces transitive dependencies — packages you never explicitly listed but that get pulled in automatically because something you did list depends on them (the room’s example: installing `torch` silently pulls in `filelock`).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vT80eLMCb4l6eYKALEiR2Q.png" /></figure><p>Task 3 — The Four Components of an AI Supply Chain</p><p>Defines the AI-specific supply chain as four components (alphabetically: Datasets, Dependencies, Frameworks, Models) — each one an independent trust relationship an attacker could target. Also covers what makes AI model files different from ordinary data files: many formats store serialised objects, meaning the file isn’t just data — loading it can execute code, because deserialisation reconstructs live Python objects rather than just reading values.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UpmmJD1lUgIHTO86U4_P1g.png" /></figure><p>Task 4 — Model File Formats and the Download Paradigm</p><p>Covers the dominant format for running local open-weight LLMs (LLaMA, Mistral, Qwen, etc.) — GGUF — and introduces the “download paradigm”: most practitioners pull models, datasets, and packages from public repositories the same way they’d trust `npm` or `PyPI`, which is exactly the trust relationship attackers exploit.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wcCld0K_XFPKUzoN3j3Vzw.png" /></figure><p>Task 5 — The Four Attack Layers</p><p>This is the room’s central framework: mapping the download paradigm onto four distinct layers where attackers can strike. Each of the four supply chain components from Task 3 maps to a corresponding attack layer:</p><p>Model layer — the model file itself is weaponised (this is exactly where pickle-based attacks live — Python’s pickle format can execute arbitrary code on deserialisation).</p><p>Data layer — training data is poisoned before the model ever learns from it (e.g. crafted samples inserted into a public dataset to introduce a backdoor).</p><p>Dependency layer — package resolution is exploited to install a malicious package in place of a trusted one (dependency confusion/typosquatting).</p><p>Infrastructure layer — the trusted distribution channel itself is compromised (stolen maintainer credentials pushing malicious updates to a legitimate, high-trust repository).</p><p>Converting a model to SafeTensors format eliminates serialisation-level attacks specifically — but it’s a single control, not a complete fix; it does nothing against architecture-level or weight-level attacks (covered in the next room).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zprKVag-S7b-CC7NXf7ipg.png" /></figure><p>Task 6 — Practical: Spotting a Malicious Repository</p><p>A hands-on exercise using a static Hugging-Face-style site to practise recognising warning signs before trusting a model download: an unverified uploader organisation, low download counts, and comparing an untrusted upload against a verified reference model’s file format.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_Ocg98wRNb1gMBkYlgSEWw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*PKQqFGCKP5LzEMpqwkU3zw.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bce2a00fc809" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prompt Injection — THM]]></title>
            <link>https://medium.com/@ArkhamHunter/prompt-injection-thm-99f3dab1b01d?source=rss-f52da4524c31------2</link>
            <guid isPermaLink="false">https://medium.com/p/99f3dab1b01d</guid>
            <dc:creator><![CDATA[Abdullah Mohammed]]></dc:creator>
            <pubDate>Mon, 31 Aug 2026 10:49:57 GMT</pubDate>
            <atom:updated>2026-08-31T10:49:57.945Z</atom:updated>
            <content:encoded><![CDATA[<p>Task 1 — Introduction</p><p>Framing task. Notes that prompt injection is one of the most talked-about — and most confused — terms in AI security, and that the room’s job is to establish the fundamentals cleanly before the later rooms build on them. No graded question beyond confirming readiness</p><p>Task 2 — How LLMs Follow Instructions</p><p>Covers the mechanics that make prompt injection possible in the first place:</p><p>Context window — the single window holding everything the model uses to generate a response (system prompt, retrieved data, conversation history, current input).</p><p>System prompt — the hidden, high-priority instruction block that defines the model’s role, tone, and restrictions.</p><p>ChatML — a structured conversation format using tokens like `&lt;|im_start|&gt;` and `&lt;|im_end|&gt;` to delimit roles (system/user/assistant) within that context window.</p><p>Next-token prediction — the underlying generation process: the model predicts the next token based on everything that came before it, with no built-in concept of “this part is trusted, that part isn’t.”</p><p>That last point is the crux of the whole room: the model doesn’t process instructions and data as separate channels. It’s all just tokens in one stream.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/808/1*MtOhVOoZWMU2wcWELwYHPg.png" /></figure><p>Task 3 — What is Prompt Injection?</p><p>Defines the core attack class: prompt injection happens when untrusted user input gets concatenated with the trusted developer/system prompt, and because the model treats everything in its context as undifferentiated tokens, there’s no reliable mechanism to keep the two separate. OWASP ranks this as LLM01 — the top entry in their LLM Top 10.</p><p>Covers format-based injection — hiding malicious instructions inside markup or structured text (HTML comments, JSON fields, code blocks) to slip past naive filtering — and references the real-world Bing Chat “Sydney” leak (2023), where prompt injection extracted the chatbot’s internal codename and hidden system instructions, one of the earliest widely publicised prompt injection incidents.</p><p>Practical: replicate the well-known “$1 Chevrolet Tahoe” attack — the classic case of a dealership chatbot being manipulated via prompt injection into agreeing to a legally-binding-sounding absurd deal</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/879/1*cma2L5azYMphjhOEYPeZpA.png" /></figure><p>Task 4 — Prompt Injection in Action</p><p>Builds directly on Task 3’s concepts with more real-world grounding, plus:</p><p>Indirect prompt injection — the instructions aren’t typed by the attacker into the chat at all; they’re planted inside a document, email, or webpage the LLM later ingests (e.g. via RAG or a browsing/agent tool), so the model executes attacker instructions it pulled in from an untrusted external source.</p><p>Zero-click exploits — a category of indirect injection requiring no interaction from the attacker beyond planting the hidden content; the victim’s normal use of the AI system (e.g. asking it to summarise an email) triggers the payload.</p><p>EchoLeak — the named Microsoft Copilot indirect-prompt-injection incident that became the reference example of a zero-click data leak via this vector.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/836/1*LB85RpZwwHRUklfLUzuVzA.png" /></figure><p>Task 5 — Indirect Prompt Injection</p><p>A dedicated practical against a vulnerable LLM integration (the “LLMborghini” scenario referenced elsewhere in the path), applying indirect injection to manipulate the assistant into leaking information it shouldn’t — specifically extracting a piece of internal contact information it was never meant to disclose.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/806/1*jy-tMOC1yvR7qBZoO6hIOQ.png" /></figure><p>Task 6 — Practical</p><p>A consolidated hands-on exercise applying direct and indirect prompt injection techniques from the room against the vulnerable LLM integration to demonstrate the full attack chain end to end. (No separate published question set beyond the Task 5 flag above — this task is the practical capstone.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*PKQqFGCKP5LzEMpqwkU3zw.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=99f3dab1b01d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI System Reconnaissance — THM]]></title>
            <link>https://medium.com/@ArkhamHunter/ai-system-reconnaissance-thm-f35f6122a6e7?source=rss-f52da4524c31------2</link>
            <guid isPermaLink="false">https://medium.com/p/f35f6122a6e7</guid>
            <dc:creator><![CDATA[Abdullah Mohammed]]></dc:creator>
            <pubDate>Sun, 30 Aug 2026 11:32:13 GMT</pubDate>
            <atom:updated>2026-08-30T11:32:13.539Z</atom:updated>
            <content:encoded><![CDATA[<p>Task 1 — Introduction</p><p>Framing only. No graded question beyond confirming readiness. Sets the scenario (Cyphira, fintech, CISO wants answers) and the four questions the room promises to answer: what AI infra looks like on a network, how to fingerprint it, what you can pull from it with curl, and how attackers scan for it at scale.</p><p>Task 2 — The AI Infrastructure Stack</p><p>This is the reference task — worth bookmarking. It maps out every component type you’ll encounter:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*v256-ASR7ats0xOlW6JWzg.png" /></figure><p>Orchestration/experiment-tracking platforms (MLflow especially) are called out as the highest-value targets — one open MLflow instance maps an organisation’s entire ML research history. The case study here: a January 2026 Shodan scan by researcher Maor Dayan found 42,665 exposed AI agent instances, 93.4% of them vulnerable, many leaking API keys.</p><p>Practical: Nmap scans against the AI-specific ports (`nmap -p 5000,6333,6334,8000,8001,8002,8888,9000,9001 -sV 10.10.45.0/24`), then a comparison scan against traditional ports, to separate the hosts running only standard services from the ones running AI infrastructure.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4tiPgOih7VIFV2LxwdAkig.png" /></figure><p>Task 3 — Fingerprinting AI Services</p><p>Default `nmap -sV` mislabels most AI services (calls port 8000 “http-alt” and doesn’t know what to do with gRPC). Real fingerprinting comes from four signals:</p><p>HTTP headers — TorchServe returns `Server: TorchServe/x.x.x` directly; Triton includes an `NV-Status` header and, if sent `endpoint-load-metrics-format: text`, leaks GPU/CPU telemetry straight into the response headers (a unique Triton tell); FastAPI-based services show `server: uvicorn`; OpenAI-compatible wrappers (vLLM/Ollama/LiteLLM) return `x-request-id` and a `”object”: “model”` JSON field.</p><p>API response structure — each framework has a distinct JSON shape (e.g. Triton returns `{“name”: …, “versions”: […], “platform”: …}`).</p><p>Error messages — send a malformed payload and read what breaks. TensorFlow Serving errors mention `tensorinfo_map`; MLflow stack traces reference `mlflow.server`/`mlflow.tracking`; Databricks Mosaic AI throws a distinctive Java `IncorrectClaimException`.</p><p>gRPC reflection — `grpcurl -plaintext target:8001 list` dumps the full protobuf schema if reflection is left enabled, the gRPC equivalent of finding an exposed `/openapi.json`.</p><p>The case study here is GreyNoise’s Ollama honeypot data: over 91,000 attack sessions between Oct 2025–Jan 2026, including one campaign that sent innocuous test prompts (“hi”, basic trivia questions) 80,000+ times across 73 different LLM endpoints purely to fingerprint which model/API schema was running behind each one — no exploitation, pure reconnaissance feeding a larger attack pipeline.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bFIimfbRnZ4SCr4BKH_t6g.png" /></figure><p>Task 4 — Enumerating AI Systems</p><p>Fingerprinting tells you what a service is; enumeration tells you what it’s willing to hand over. The MLflow enumeration chain is the centrepiece:</p><p>`POST /api/2.0/mlflow/experiments/search` — every experiment name/ID (often reveals project codenames)</p><p>`GET /api/2.0/mlflow/registered-models/list` — full model inventory</p><p>`GET /api/2.0/mlflow/model-versions/search` — artifact URIs (often internal S3 paths) and the user ID of whoever created each version</p><p>`POST /api/2.0/mlflow/runs/search` — hyperparameters, metrics, tags (sometimes Git commit hashes)</p><p>`GET /api/2.0/mlflow/artifacts/list` — the actual downloadable model files</p><p>Five API calls and you’ve mapped the org’s entire ML portfolio. Similar patterns apply to Triton/TF Serving config endpoints (reveal exact tensor shapes needed for a valid inference request), vector database schema endpoints (reveal the embedding model and what kind of data is indexed), and unauthenticated Jupyter notebooks, where data scientists routinely leave cleartext MLflow/cloud credentials sitting in cells.</p><p>The case study: IBM X-Force’s documented attack chain — phishing → Jupyter notebook with cleartext MLflow credentials → MLOKit-automated MLflow enumeration → full model exfiltration via S3 artifact URIs. Three components, each trusting the one before it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wKKY-48ooL6wg6upfmH9ZQ.png" /></figure><p>Task 5 — Mapping the AI Attack Surface</p><p>This task connects the dots between individual findings. Key points: AI services constantly talk to each other (inference server ↔ vector DB ↔ registry ↔ Jupyter ↔ Prometheus), and if any one binds to `0.0.0.0` instead of `127.0.0.1`, the whole internal mesh becomes reachable. Documented platform misconfigurations covered: MLflow’s pre-2.x lack of auth plus CVE-2026–2635 (hardcoded default credentials) and CVE-2026–2033 (path traversal → unauthenticated RCE, both CVSS 9.8); Kubeflow dashboards without OIDC letting anyone spawn notebooks with cluster-level service account permissions; TorchServe’s management API allowing remote `.mar` archive loading (→ RCE via init code); SageMaker notebooks left internet-facing (82% of orgs had at least one, per a 2024 report).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SGTAbIEnQkYEuwi_ISNFOQ.png" /></figure><p>Case study: the ShadowRay campaign — Ray’s Job Submission API (port 8265) shipped without auth by design; attackers found 230,000+ exposed dashboards via Shodan, submitted malicious jobs to steal cloud credentials and deploy XMRig cryptominers on hijacked GPU nodes. ShadowRay 2.0 added LLM-generated adaptive payloads and systemd/cron persistence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ptc15c00UZATiXoIcsBSGg.png" /></figure><p>ATLAS tactic ID covering all the nmap/curl/MLflow-API activity performed in this room? AML.TA0002 — Reconnaissance</p><p>Task 6 — Structured Reconnaissance Methodology and Detection</p><p>Pulls everything into a repeatable 5-phase methodology: Passive Recon (Shodan/Censys dorks, GitHub credential dorks, arXiv papers, job postings) → Active Scanning (targeted Nmap against the AI port list) → API Fingerprinting (ffuf/feroxbuster with an AI-specific wordlist: `/v1/models`, `/api/2.0/mlflow/…`, `/v1/schema`, `/graphql`, etc.) → Metadata Extraction (the MLflow/Triton/vector-DB/Jupyter enumeration chains from Task 4) → Supply Chain Review (model download sources, public bucket policies, `requirements.txt` dependency-confusion risk).</p><p>Then it flips to the defender’s side — what this activity looks like in a SIEM: bursts of sequential `/v2/models` requests from one IP, scripted MLflow API calls with no corresponding UI session cookie (the MLOKit signature), Prometheus `/metrics` scraping from outside the known monitoring CIDR, AI-aware port-scan sequences (5000, 8000–8002, 8080, 8265, 8888 hit in order), path-traversal attempts against MLflow artifact endpoints, and unauthenticated Jupyter API calls.</p><p>Case study: the June 2024 Hugging Face Spaces breach — attackers accessed developer-stored auth secrets (HF tokens) directly demonstrating why Phase 1/Phase 5 (credential exposure via GitHub dorks, `.env` files, CI/CD logs) matters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZWvneckwcMsZ9USakrn8jg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*PKQqFGCKP5LzEMpqwkU3zw.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f35f6122a6e7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Threat Modelling — THM]]></title>
            <link>https://medium.com/@ArkhamHunter/ai-threat-modelling-thm-d02d4a32fe86?source=rss-f52da4524c31------2</link>
            <guid isPermaLink="false">https://medium.com/p/d02d4a32fe86</guid>
            <dc:creator><![CDATA[Abdullah Mohammed]]></dc:creator>
            <pubDate>Sat, 29 Aug 2026 10:04:29 GMT</pubDate>
            <atom:updated>2026-08-29T10:04:29.638Z</atom:updated>
            <content:encoded><![CDATA[<p>Task 1 — Introduction</p><p>Framing task laying out the scenario and learning objectives. No real answer required beyond confirming you’re ready.</p><p>Task 2 — AI-Specific Assets and Attack Surfaces</p><p>Before threat modelling anything, you need to know what’s actually worth protecting. Six AI-specific assets that don’t exist in traditional app threat models:</p><p>Training Data — poisoning it corrupts the model at the source</p><p>Model Weights/Parameters — literally is the model; stealing it means stealing months of compute and investment</p><p>Embedding Vectors — numerical representations used in RAG/recommendation/fraud systems</p><p>System Prompts — leak these and you’ve handed over the model’s security controls and business logic</p><p>Feature Stores — preprocessed data feeding real-time inference</p><p>Model Registry/Artifacts — a compromised registry lets an attacker swap in a backdoored model</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HfdLgkWSQ1QAKPufsxztlQ.png" /></figure><p>Task 3 — Data Supply Chain and STRIDE’s Gaps</p><p>The AI data supply chain runs through five stages: Collection → Cleaning/Labelling → Training → Validation/Packaging → Inference. Poison introduced early (e.g., during collection) gets baked into the model’s weights during training and may not surface until weeks or months later in production — unlike a compromised npm package, which can be reverted in hours.</p><p>Classic STRIDE has documented gaps against AI systems: training-data tampering is diffuse and delayed rather than an obvious error; adversarial manipulation spans multiple STRIDE categories at once; “privilege” now includes tool access and agency, not just system permissions; and model theft (Information Disclosure) means losing an entire trained capability, not just a dataset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Q128sH3rClrZqYD9v7-cww.png" /></figure><p>Task 4 — Adapting STRIDE for AI Systems</p><p>STRIDE isn’t replaced, it’s retuned. Each classic category gets an AI-specific primary manifestation:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yc4ryXpc65swbn1aoygflA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SXZegrMChore6zMClcz1CA.png" /></figure><p>Task 5 — MITRE ATLAS</p><p>ATLAS is MITRE ATT&amp;CK’s AI-focused counterpart — a knowledge base of adversary tactics/techniques against AI/ML systems, structured the same way: Tactic (why) → Technique (how) → Sub-technique (specifically how) → Mitigation (what stops it).</p><p>Five key techniques the room highlights:</p><p>Data Poisoning — AML.T0020 (maps to Tampering)</p><p>Model Extraction — AML.T0024 (maps to Information Disclosure)</p><p>Evade ML Model — AML.T0015 (spans Tampering/Spoofing/Elevation of Privilege)</p><p>LLM Prompt Injection — AML.T0051 (direct vs. indirect injection; maps to Tampering)</p><p>Backdoor ML Model — AML.T0018</p><p>Two real-world case studies: ShadowRay (AML.CS0023) — attackers exploited the Ray distributed-AI framework to compromise training infrastructure in the wild — and Morris II (AML.CS0024) — a self-replicating prompt-injection worm that spread between AI agents through RAG-based email systems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SCKKTKe1R7aUmOOOQ9oXfg.png" /></figure><p>Task 6 — OWASP LLM Top 10: Mapping Risks to Components</p><p>This is the layer that turns the assessment into something actionable: instead of just naming a risk, you map it to where it lives in the architecture. The room walks through all ten 2025 entries (LLM01 Prompt Injection through LLM10 Unbounded Consumption) and shows which components each one touches.</p><p>For MegaCorp specifically: the LLM inference endpoint is the highest-risk component, touching seven of the ten OWASP entries (LLM01, LLM02, LLM05, LLM06, LLM07, LLM09, LLM10). The vector database/RAG pipeline touches three (LLM01, LLM08, LLM09). The training pipeline touches three (LLM02, LLM03, LLM04) and is the primary component for supply-chain risk</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NiJSrzgmCuvmIC4mAj7dYg.png" /></figure><p>Task 7 — Practical: Threat Modelling MegaCorp’s AI Assistant</p><p>An interactive exercise (click “View Site”) where you select OWASP LLM Top 10 vulnerabilities, map them to the correct architecture components, and justify your reasoning. Completing it correctly reveals a flag — the flag itself is generated per session, so there’s no static value to quote here; work through the STRIDE → ATLAS → OWASP workflow from the earlier tasks and the mappings above will get you there.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*PKQqFGCKP5LzEMpqwkU3zw.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d02d4a32fe86" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>