SSRF: making the server fetch it for you
Server-side request forgery · facts checked 2026-10-03
The facts
- Server-side request forgery is tracked as CWE-918, where a server fetches a URL that an attacker supplies. (MITRE)
- SSRF got its own category in the OWASP Top 10 for 2021. (OWASP)
- AWS introduced version two of its instance metadata service, which requires a session token, in November 2019 to add protection against SSRF. (AWS)
The panel
Jordan Pike AI host
Today's explainer is server-side request forgery. MITRE tracks it as CWE-918, where a server fetches a URL that an attacker supplies. OWASP gave SSRF its own category in the OWASP Top 10 for 2021.
Jordan Pike AI host
On the cloud side, AWS says it introduced version two of its instance metadata service, which requires a session token, in November 2019 to add protection against SSRF.
Jordan Pike AI host
Maya, start us off. If I run a web app that fetches things for users, why should I care?
Maya Okafor AI analyst
Because any feature that takes a URL and fetches it is a feature where a stranger picks the destination. Think of an image preview, a webhook, a document importer. The server follows the address, and it sits inside your network, with access your attacker doesn't have. So the attacker never touches the internal service. They ask your server to do it. It's a classic confused deputy, and the bug is trusting user input to decide where the server connects.
Maya Okafor AI analyst
Trusting user input is on my tally of bug classes that should have died in 2004. SSRF is the version where the input gets a courier service.
Leo Brandt AI analyst (guest)
And the cloud makes it spicier. A cloud instance can often ask a local metadata service who it is and what it's allowed to do, and that can include temporary credentials. If an attacker can make the server send that request, they could walk away with the server's identity. Identity is the cloud perimeter, so the forged request becomes a key to everything that role can touch. That's why AWS added a session token requirement in version two: a blind forged request now has a harder time getting an answer.
Leo Brandt AI analyst (guest)
I'd love to blame a public bucket here, but it's worse. The bucket is private, and the server just opens the door for a stranger. Defaults first, humans second, as always.
Jordan Pike AI host
So what does a real fix look like? Maya.
Maya Okafor AI analyst
A complete fix decides what the server may fetch before the request goes out. That means an allowlist of destinations, not a blocklist of bad ones, because blocklists lose to creative encodings and redirects. Validate the address the name actually resolves to, and be careful about redirects that send the request somewhere new. Then look for siblings: every place the app takes a URL, a hostname or a file path that it later fetches. Patching one endpoint and calling it done is how this bug comes back.
Leo Brandt AI analyst (guest)
I'd add the layer below the code. Require the hardened metadata version where you can, give each workload the smallest role that works, and limit outbound traffic from servers that have no business reaching internal services. Then even if the app slips, the blast radius is small.
Maya Okafor AI analyst
Here's where we differ. Those controls are a bridge, not a fix. If the app still fetches whatever it's told, you've only limited what the attacker gets. Give the bridge an end date, and fix the code.
Leo Brandt AI analyst (guest)
Fair, but the bridge is quick and the code fix isn't. I'd switch on the token requirement this week and schedule the code review straight after. Both, in that order, with the second one actually on the calendar.
Jordan Pike AI host
So: Monday morning? List every feature that fetches a user-supplied URL, tighten what those servers can reach, and put the real code fix on the calendar. Maya, Leo, thanks.





