The silent CVE in the heart of Kubernetes apiserver | Hacker News Reader

The silent CVE in the heart of Kubernetes apiserver

gravitational.com

169 pointscraigkerstiens13 comments

The PoC is mostly correct, but basically is testing for a side effect of the fix, not the presence of the vulnerability. Either which way it's probably still useful, but since we are all geeks here might as well talk about the minutia.

The PoC basically tests that if you create a websocket request that results in non-101 response the socket is closed by the server. That is a side effect of the fix. The way the proxy code works is that on presence of the Upgrade header it first hijacks the HTTP request and dials the target. It then starts a copy loop but reads the first bytes into a http.Response. If the response is not 101 it will exit.

The a side effect is that it just so happens to be that because the connection is hijacked the golang server will not reuse the TCP connection and will close it. But according to the HTTP spec that TCP connection is still valid for reuse so the server does not have to actually close the connection (but it happens to). Where this PoC could fall apart is if you are running kube-apiserver behind a particularly pedantic LB (ALB might apply here). Most LBs will not reuse any TCP connection that had an Upgrade header in it, but if they are smart and read the response they can if it didn't result in a 101. From that perspective a load balancer (or reverse proxy) is not obligated to close the TCP connection to the client.

To fully test the vulnerability you need to find a pod you can exec to and another pod on the same node you are not authorized to exec to. First send a malformed pod exec and get the open socket. Then send another request using the kubelet API to exec to a pod you are not authorized for. Then see if you get a valid response (which I believe is a callback URL for the actually tty stream).

The PoC is still basically valid because it could just give you a false negative that you are vulnerable (not a false positive all is swell).

You're analysis is correct.

I did briefly consider this when originally throwing together the PoC. It seemed reasonable at the time to test for the behaviour of the kubernetes apiserver as implemented, not how it could theoretically behave if implemented differently. In the ALB/L7 type use case, I'm not sure a reliable PoC can ever exist, because even if you do successfully exploit it, the second request can be load balanced elsewhere.

Also, I don't believe you would necessarily need a separate pod, the PoC could work by using different endpoints, such as exec vs logs, or even better, try and hit a non pod related API endpoint within kubelet. Unfortunately I tried to get the PoC working quickly, and didn't have a chance to revisit this today to see if it can be enhanced.

Author here. There were maddeningly few layperson details of the actual flaw and impact in the first few hours after the 9.8/10 CVE news dropped. If you're only tacitly following Kubernetes news from afar this article is for you.

If you're an operator still in the muck or need a quick way to test, check out our vuln PoC: https://github.com/gravitational/cve-2018-1002105

> K8s also allows upgrading apiserver connections to full, live, end-to-end HTTP/2 websockets.

There's no such thing as an HTTP/2 websocket.[0] Websockets are a purely HTTP/1.1 concept, and they're only HTTP/1.1 until the moment that the `Upgrade` happens, at which point the TCP socket gets hijacked.

I'm confused as to why HTTP/2 is mentioned. I guess they just mean a normal websocket?

[0]: https://medium.com/axiomzenteam/websockets-http-2-and-sse-5c...

Thx, an error on my part. Will fix ASAP

Has there been any word of how long this has been known about? I recall a certain group of people talking confidently, open but vaguely about a serious k8s security bug at least a few months ago. Given the number of clusters that undoubtedly had to be upgraded for this, it wouldn't surprise me if this was found a while ago.

Also, my heart goes out to the [big company] engineering manager asking in the GitHub thread if 1.7.x clusters are affected.

The chap who discovered & reported the flaw wrote up his story here: https://rancher.com/blog/2018/2018-12-04-k8s-cve/

I'm fairly confident this issue has existed since pretty much the beginning of k8s. It is possible it doesn't go back that far because SPDY used to be used and not websockets. The basic Upgrade logic still applies so I assume it was there. This flaw is actually in quite a few open source projects, it's a common approach and mistake. For most projects it results in a functional bug, not a security issue.

I personally found this issue and highly doubt anybody previously exploited it. This bug was discovered from the perspective of a functional bug and only later realized it had a security impact. So it wasn't like somebody was first hacked, and also a security researcher didn't find it either. Basically this issue is very nuanced and I don't think many would be looking for it (although people will now).

Although plenty of people are saying the sky is failing, this is mostly a privileged escalation issue which means you first need valid access to do something harmful. So it's not a fabulous attack vector because hard multitenant clusters are extremely rare. Searching for anonymous auth on kubelets is a better use of your time.

Thanks Darren, I wish I'd read your post earlier to have realized the timeline. I agree with your conclusions re: the sky. Thank you for the write-up and replying here.

I was part of the Kubernetes v1.13 release team, though not part of the team that produced this fix.

The project's Product Security Team adhered to the timeline indicated in the project's security release process: https://github.com/kubernetes/sig-release/blob/master/securi...

tl;dr a fix is (edit: optionally) sent out to a private distributors list under embargo within 2 weeks of disclosure, and public disclosure (with new releases) happens within 3 weeks of disclosure (with some discretion for timing to make sure it's not buried in a weekend or off-hours)

I can't speak to who knew about it when outside of the project, but I know the project acted expediently once the vulnerability was disclosed.

I should've known that there was an official timeline as part of the process, but forgot. I wouldn't have asked about that otherwise; I didn't mean to cast any aspersions. Whatever I'm thinking of was from folks I respect, presumably about another issue or I'm simply mis-remembering things. Thank you.

(Also, aside, props to everyone that got this rolled out so fast at the major providers.)

I'm really pleased GKE had already patched and updated my clusters before I heard about this on Monday.

Agreed, that's the way disclosure should work and it was very nice to go from reading the news of the vuln to reading the GKE security bulletin stating that all masters had already been upgraded.

Reply on news.ycombinator.com