SSRF in AI Table Generation via uploadUrl: raw fetch without blacklist protection
7.7
High
CVSS 3.1 base
What to do
If you run budibase 3.41.0 and earlier, upgrade to a release after 3.41.0.
Anyone who can reach it over the network, with a low-privilege account and no user interaction, can read sensitive data, and the impact reaches beyond the vulnerable component.Read from the CVSS 3.1 vector
To exploit it
What it gets
My analysis of this finding: root cause, the vulnerable code, the impact and the fix, in the site’s format. The official advisory of record is linked above.
| Package | Ecosystem | Vulnerable | Fixed in |
|---|---|---|---|
budibase | npm | <= 3.41.0 | > 3.41.0 |
Budibase builds attachment columns during AI table generation by downloading whatever URLs the model emits. The download path, uploadUrl in packages/server/src/utilities/fileUtils.ts, calls node-fetch directly with no SSRF guard, while every other outbound request in the codebase goes through fetchWithBlacklist. A builder-level user who steers the LLM toward an internal URL turns that gap into a full read server side request forgery, because the fetched body is stored and the storage URL handed back.
The sink is a bare fetch on an attacker influenced url, packages/server/src/utilities/fileUtils.ts:23:
const res = await fetch(url) // raw fetch, no blacklist checkThe bugfetches the model supplied url with no blacklist check
This is reached from POST /api/ai/tables. When the LLM generates table data that includes an attachment column, the URL in that column is passed straight into uploadUrl. The rest of the server routes outbound traffic through fetchWithBlacklist from @budibase/backend-core, which rejects private and link local IP ranges. uploadUrl is the one place that skips it, so the private range protection that exists everywhere else simply is not applied here.
The request body of whatever URL uploadUrl retrieves is written to object storage, and the resulting S3 style URL is returned to the caller. That makes this a full read SSRF rather than a blind pingback: an attacker does not just cause a request, they get the response back.
A builder influences the prompt or the generated table content so that an attachment URL points at an internal target, for example the cloud metadata service at 169.254.169.254 (AWS IMDSv1), an internal admin panel, or another service that is only reachable from the server’s network position. uploadUrl fetches it, stores the response, and returns a link the attacker can read. The practical payoff is reading cloud instance metadata and credentials, reaching internal only HTTP services, and exfiltrating responses that should never have left the trust boundary.
I want to be precise about reach, because it bounds who is actually affected. In my notes this code path was present on the master branch and I did not confirm it in the v2.8.x release tags. The honest status is that the vulnerable fetch(url) and its /api/ai/tables trigger are real in the tree I reviewed, and the open question at the time of writing was which 3.x release first shipped it. That gate matters: code that only ever lived on a development branch has no installed users to harm.
Route this call through the same control everything else uses. Replace the raw fetch(url) in uploadUrl with fetchWithBlacklist so private, loopback and link local destinations are rejected before the request is made, and validate the URL scheme and host before download. There is no reason for this one download helper to bypass the blacklist the rest of the server already enforces.
Sanaan Fayaz Wani (GitHub sfwani) reported this vulnerability to the budibase maintainers under coordinated disclosure and is credited as a reporter in GHSA-3c52-v5v2-3r56, published 2026-09-17.
All published findings: advisory index.