About a year ago I told you about the 100. 000$ GCP Prize from the google bug bounty program, awarded for the best vulnerability found in the Google Cloud Platform. Now a year later, it’s time again! And the prize was higher too - $133,337. In this video I will tell you the story and technical details of a new amazing winning bug. It’s basically a Server Side Request Forgery attack, but with the impact of a remote code execution inside of Google.
So let’s meet this year’s winner from Uruguay! “So my name is Ezequiel Pereira. ” This is him on our call when we talked about his bug, but at the time, he didn’t know that he won the 133. 000$ yet. I pretended to just be curious about his blog post: “This sounds really insane and I was wondering if you would be interested in making a video together. ”. He said yes, and so he started explaining his bug.
00:57 - “They classified it as RCE, because an attacker could potentially execute code by calling some internal google endpoints. And those internal google endpoints, seeing that the request comes from the cloud deployment manager, might allow the attacker to do actions that an external user shouldn’t be allowed to do. So it is equivalent of RCE. I never got to the point of actually executing code on google, especially because they cut me off. Because they were treating this as an incident” So his bug allowed him to basically perform arbitrary requests inside of the trusted Google network.
And he could potentially reach critical endpoints. We will soon talk about this internal network, because it’s really fascinating - this bug combines so much knowledge about google internals. But as he said, he never got so far to exploit it further. Google knows what impact this special server side request forgery has, and classified it as a critical remote code execution.
02:06 - “If you get into the internal network, and are able to issue requests, it’s like RCE” It’s not true in any SSRF case. Because internally in Google, requests between services have to be authenticated. But in this case the source of the request was authenticated with apparently high privileges, thus in this case, being able to issue requests, is like RCE That’s why they also treated it as an incident. This is a bug that needs to be investigated because maybe somebody else, an actual attacker used this.
Of course Ezequiel also got awarded the regular bug bounty awarded when he reported it, so how much was it? “You have a rewards table, and you can see they always pay 31k for RCE issues in their main google products. ” What Ezequiel didn’t know was, that I had the Google VRP team on standby, to join the call at any moment. I acted surprised why somebody would want to join the call. I accepted them all. “We concluded that you won the top prize. Which is a $133,337.
We wanted to surprise you in person. ” “Woohoo. ” “Thank you. Thank you very much. ” Congratulations! “Well I didn’t really expect to win. I thought like maybe the firebase bug, that like sent tons of notifications would win” So I guess Ezquiel needs to update his blog post with the facts! It’s not just 31k. It’s a total of 164. 674$. Damn!!!! But now I hope you are ready to hear the story about this bug.
04:16 - “Google Cloud is huge. I can assure you that I haven’t even touched like 70% of it. 70% of google cloud is completely unknown for me. I’m no google cloud expert. I just go through the documentation. And if I kinda understand something I begin looking into it. I researched App Engine a lot, because it’s easy to use, easy to understand. And it is really interesting” App Engine is one of the big main products of Google Cloud. Basically it allows you to host web applications.
04:56 - “So while looking at App Engine flexible environment, I stumbled upon Deployment Manager. So here for instance I deployed something on App Engine flexible environment, and you can see calls to the Deployment Manager API. And you can see like the user is gae api prod. google. com service accounts. Well, I decided to look into it because if it was being used by App Engine, it will probably be used by other GCP products. And when something is used internally at Google, even if it is a public issue like deployment manager, they sometimes hide internal settings, internal stuff, that if they are not well protected they could be exploited by an attacker.
And that’s what happened here. ” So the “Cloud Deployment Manager” is one of those many products that he didn’t know. But then saw in the logs, that App Engine uses it internally to manage the servers. And I think it’s very clever how he thinks about google products that they also use internally.
06:10 - “For instance I think every, or almost every, withgoogle. com website is really an app engine application. And sometimes you can see that the website is integrated with the other google services. So you begin to wonder, if it is running on App Engine, how is it connected to internal google stuff. Or for instance they use GCE, Compute Engine a lot too. So while using them, sometimes they need to do internal stuff. So sometimes they build into the public tool internal stuff, that they just hide somehow from the public.
Because they are only meant for internal users. If you pay attention sometimes documentation also references internal stuff. And you say, what does this mean in the documentation?! And it doesn’t make sense, because well, it is intended only for googlers. ” I think this is a very valuable tip for aspiring google bug hunters and probably the most important takeaway of this video. Approaching the target with the mindset, that if a product is also used internally by google, maybe there are undocumented internal features exposed that could be exploited.
And he saw that the Deployment Manager was used by App Engine. So he decided to hunt for bugs there.
07:43 - “So I knew nothing about Deployment Manager. I didn’t understand what it was for. Even right now I am not pretty sure why it exist or how it is useful for a developer. But yeah, I had to read the deployment manager documentation like 4-5 times until I kind of got the idea what it was for. ” Let me try to give you a brief summary of what the Deployment Manager is. It all starts with a configuration or template describing some resource. Here for example a compute-instance, so a basic server, being deployed in a US datacenter.
It also has a harddrive attached, with a debian image on it, and it also has an external network interface with external NAT. So this is a whole machine description. Now you can take that, and send this to the Deployment Manager, and it will then setup this server for you. So you can kinda imagine this like a docker-compose file, or maybe a Kubernetes Deployment object. This is just a Google Cloud Deployment Manager config. Now let’s think about App Engine, which is used to host web applications.
When you deploy an app, it seems like that App Engine uses the Deployment Manager, to describe the server where that app will be running.
09:00 - “I began playing like creating my own templates. And creating resources. And Looking how it works. Looking at the different features. For instance you know there are two public versions of deployment manager. You have the V2 version and v2beta version. So I looked at the difference of the two versions. ” And one of those differences are TypeProviders. It’s a very confusing name, but it’s important to understand. So in a Deployment Manager config file you have a type field, and that describes what kind of resource or server you want.
In this case it’s a compute. v1. instance. But that’s something Google Cloud specific. What if your company uses, besides Google Cloud, also your own datacenter with machines you want to manage too? TypeProviders can be used for that.
09:48 - “A type provider exposes all of the resources of a third-party API to Deployment Manager that you can use in your configurations. These types must be directly served by a RESTful API that supports Create, Read, Update, and Delete (CRUD). ” So as long as you can provide a simple HTTP REST API for your datacenter, that implements Create a Server or delete a Server, then you can define a TypeProvider describing your API, and then you can use Deployment Manager, referencing your own type, to talk to your datacenter’s API.
This way you can manage all your cloud resources in one place.
10:24 - Anyway. Here is an example API request to create your own type provider with the v2beta version. And most important is here the descriptorURL. It points to a JSON file. And this JSON file is like a swagger API definition. This is what actually describes your API endpoints where you implement your create or delete resources stuff. The options field is also interesting. You can see here that it defines an Authorization header. It’s obviously important that when you implement your own API to manage your servers, that the API has some form of authentication.
And in this example you send a google oauth token along, that you can then check. But now let’s send this request.
11:04 - “So you can see the operation was completed. So the type provider was created. And if I go to my server I can see that this IP connected to my HTTP server and retrieved the descriptor document for my fake API. and it provided the access token, I told it should provide. ” And now maybe you can already see where this is going. The bug Ezequiel found is a server-side-request-forgery attack. And here we control a URL that the Deployment Manager sends a request to.
So is it as simple as for example pointing this at localhost, or some other internal IPs or hostnames? “If I try to create a type provider that talks to an internal server, like server side request forgery, it will try at first to create the type provider. But it will fail. It will say error processing request. Error fetching URL localhost. Error excluded ip. It won’t let me do internal requests just like that. I tried like setting my own domain that will like point to an internal service.
You see that here it failed on the creation of the type provider, so I also tried like setting my own domain that at first points to a valid service, and once the type provider got created It tried changing my domain to an internal server to see Maybe I could bypass it that way. I’m not an expert on all of this. So maybe someone looks at this, tries something and finds a way to get SSRF through here. But I was not able to. ” Huh. That would have been too easy, right?! There is more to come.
As a mental exercise, try to think about what you would try next. Or just try to guess where this is going. This is what I do trying to figure out if I could have found this bug too.
13:08 - “then I moved on, and some days later I decided. Ok. Maybe I can find an internal method used by the Deployment Manager. Because remember. Google when using this public tools internally, sometimes they hide internal stuff inside. And sometimes they are internal hidden methods in the api. So I know a way to list all the API methods, even undocumented ones. And it is through the metrics page here in the cloud console. And funnily enough it doesn’t only show the public methods.
But also some internal ones. If there are. So here you can see for instance the GET operation method in the v2 version. But looking at this, I noticed that you have v2 and v2 beta. But also here for instance here you have dogfood and you have alpha version. And those versions are not documented publicly. And I said, Ok I’m going to look into them. ” Mindblowing. Like a detective finding small puzzle pieces. So let’s see what happens when you try to send requests to those different versions.
14:27 - Btw. look at my face during the call. I’m in total awe right now.
14:32 - “I can get an operation on the v2beta version. I can also do it on the v2 version. Let’s see what happens on the alpha version. Can I call a method on the alpha version? Yes I can. Can I call a method on the dogfood version. Yes I can. If I try a version that doesn’t exist. No it doesn’t default to a public version. It just says not found. Now I know alpha and dogfood are real versions”.
15:02 - And every google bughunter should get excited when they read dogfood. Here is a googleblog about testing from 2014 describing their concept of Dogfooding: “Google makes heavy use of its own products. Because we use them on a daily basis, we can dogfood releases company-wide before launching to the public. These dogfood versions often have features unavailable to the public but may be less stable. ” Now it’s not necessarily a security issue that you can access a dogfood version publicly, but if it’s a less stable test version, with maybe bugs, there is a higher chance for it to have security relevant bugs too.
So it totally makes sense to now go after this dogfood version of the API and see if there are new features that are not in the public release, that could be exploited.
15:53 - “So I begin looking into the requests. This method is called list types. it tells you the built in type providers of deployment manager. So for instance here at first you can see that deployment manager is able to manage spanner instances. I was looking into this. I scrolled here. Said okay, all of this sounds like stuff that is already documented. Until I got here! I was looking at the builtin types and suddenly I found with the dogfood version there is something with googleOptions.
This is not documented. This is not in the public versions. So I was wondering what is it doing here. There I found one difference. There I found one difference with the public api. yah I looked into it and said, if it is on the builtin type providers of deployment managers, maybe I can set it also on my own type providers” Ezequiel maybe just found an undocumented internal googleOptions field and was wondering if he can set it on his own Type Provider, and maybe it does something interesting.
17:07 - “As soon as I found this, I was really interested in it. Especially because I saw this. GSLB target. And I know that GSLB is the internal Global Service Load Balancer of Google. And if you read the SRE book you can see that it might let you send requests to internal servers. ” The Google SRE book he mentioned is really cool. It has been on my reading list for MANY MANY years. But I cannot read books, so I never did. Though even though I haven’t read it, I know it’s amazing.
Because as Ezequiel just said, you can learn about some cool internal Google stuff. And in this case, the Global Service Load Balancer (GSLB) is important. In the chapter about “The Production Environment at Google” you can read “Our Global Software Load Balancer (GSLB) performs load balancing on three levels. Frontend, services and internal remote procedure calls. The frontend handles your typical DNS queries for domains like google. com. But internally Google uses their own system.
Service owners (so basically developers) specify a symbolic name for a service, a list of BNS addresses of servers [… ]. GSLB then directs traffic to the BNS addresses. ” So internally google uses BNS addresses to identify servers. Further down we get an example of how an HTTP request to a google service is handled.
18:37 - “first, the user points their browser to shakespeare. google. com. To obtain the corresponding IP address, the user’s device resolves the address with its DNS server. This request ultimately ends up at Google’s DNS server, which talks (internally) to GSLB. As GSLB keeps track of traffic load among frontend servers across regions, it picks which server IP address to send to this user. The browser connects to the HTTP server on this IP. This server (named the Google Frontend, or GFE) is a reverse proxy that terminates the TCP connection.
The GFE looks up which service is required (web search, maps, or—in this case—Shakespeare). Again using GSLB, the server finds an available Shakespeare frontend server, and sends that server an RPC containing the HTTP request. [… ] the frontend server contacts GSLB to obtain the BNS address of a suitable and unloaded backend server. ” So as you can see the Global Software Load Balancer is inside the google network. That means, if you somehow can send requests to a service with their BNS address, you are really deep inside of Google.
And you might be able to send really critical requests to very important internal servers.
19:53 - And now coming back, here you have a GslbTarget field which sounds like you can maybe specify a BNS address. Those addresses are not necessarily secret, but they are also not really public. Though sometimes they “leak” out. You can find some of them appear in logs or API responses. But Ezequiel also had this funny story to share.
20:15 - “There is this screenshot that I wanted to show you. This screenshot is from a website. An internal google website. Not internal. Uhm. it was a webpage that was exposed publicly by mistake. Until yesterday. Yesterday they blocked access to it. But here you can see that you have GSLB addresses. This is an example of how someone might find GSLB addresses just by luck. ” So fascinating right? All those puzzle pieces slowly coming together. And we are slowly getting to the vulnerability.
So we just found this GslbTarget value, and there is this idea, maybe the domain you specified here is overwritten by the GslbTarget, and so you can use this request from the Deployment Manager to send requests internally.
21:16 - “So for instance here I’m trying to breach the corporate issue tracker api. So this is the symbolic name for the issue tracker api. the issue tracker api is issuetracker. corp. googleapis. com. So usually you don’t have access to that. Okay I’ll try specifying this on the gslbTarget. So these values, I set this to true and this false. Just because through trial and error I saw those values worked. So I set them to that. And Transport again, I saw harpoon being a transport value.
So I couldn’t get SSRF this way, because as you can see, my server got hit by the deployment manager request. But my server is not inside gslb. So the request did not go through the internal load balancer. So I didn’t get SSRF. ” Good idea. But it didn’t work. Would you have given up at this point? I might have. But Ezequiel had another idea.
22:29 - “I suspect it has to be with Transport here. Because well it doesn’t make sense for the other values to have anything with going through GSLB or not. But I didn’t know what values to put here. Should I put “internal”? I tried bruteforcing it, but I just didn’t get it right. ” Okay maybe Eziquel just has to find the correct transport method, and then Deployment Manager will honor the gslbTarget address and send a request to this internal service. The problem is, this is an enum field.
So you need to exactly know the name of the correct transport. That’s why he tried to bruteforce it.
23:15 - “But at the time I didn’t know what value it could be. And I tried bruteforcing it. Like internal, or corp, or whatever, And I couldn’t get it working. I couldn’t get SSRF. It would always keep my own server and not the internal google server that I wanted. And once you know what value goes there it’s really obvious. But at the time it wasn’t obvious for me. So I went through like, I spent weeks stuck here. And one day finally I got an idea. I wanted to use protocol buffers.
” Wait for it. This is so smart! Protocol buffers, or protobuf, is a binary serialization format for data. It’s like JSON, just not readable by humans. It’s binary data. JSON is what normal people on the internet use. And protobuf is like what cool people use. And it’s from google. Thus tons of google services use, or at least support, protobuf instead of JSON. Or to be more precise, many Google APIs do not only support your basic boring boomer HTTP/1.
1, with your neat easily readable HTTP headers. But they also support gRPC. Which is a protocol using protobuf over HTTP/2. And HTTP/2 is also binary. I wonder how many of you have actually heard or worked with HTTP/2 before.
24:41 - “So if you know protocol buffers, they are a serialization format that google uses a lot. And in that format enumerations are encoded to binary as numbers. Instead of strings. So for instanced here instead of specifying HARPOON or OAUTH or GOOGLE, I would just have to specify their enumeration number. ” This is so clever! In your human readable JSON, you would have to know exactly the name of the enum. But in binary protocol buffers, data is tightly packed, and for an enum with just a few options, you don’t want to waste space and store long strings.
Because protocol buffer definitions are compiled and shared between client and server, you can just encode it as numbers. AND THEN YOU WOULDN’T HAVE TO KNOW the exact name for the transport. You can just try out all of them. And so Ezequiel tried to interact with this API via gRPC. Or actually he tried to use a trick to more easily work with it. Google sometimes supports protobuf over HTTP/1, instead of HTTP/2.
25:48 - “One way of looking that, is like for instance specifying alt parameter here I want proto. But it says, proto over http is not allowed for this service, in this case deployment manager. So in this case google disallows this fallback. So I cannot use protocol buffers on deployment manager. Or that’s what I thought at first. Another thing google has, that their staging environment are very often accessible. So I said, okay if I cannot use proto over http on the production environment of deployment manager.
Can I somehow maybe access the staging environment? So through experience again I know that in some APIs to access the staging environment you just need to prepend staging_ in the version name. Just like that. Okay I did this. I called the staging environment, and it worked. I invoked a method on the staging environment. It says my type provider does not exist, because of course it exists on production not staging. So let’s create it on staging” Oh my gosh he found another hidden version! “So here, look.
In the staging environment of deployment manager, I am able to use proto over HTTP. and well this is binary i don’t understand. there is another way to get a protobuf response that is through the content type. Content-type application x protobuf. And again, the same.
27:43 - But wait, it gets more crazy. “Something else I wanted to mention is that Google APIs don’t only serve from googleapis. com, they also sometimes serve from client6. google. com. But if you try to get a protocol buffer response from google. com, it will freak out. Because it says, request unsafe for trusted domain. I mention this because the way google bypasses this when they need to use client6. google. com, is through a header X-Goog-encode-response-if-executable=base64.
So i run this. And I have the response in protocol buffers encoded in base64. If you don’t have the protobuf definition of a protocol buffer message, luckily the protoc the protocol buffers compiler tool, has an option to decode the raw message and just give you the field numbers. And here I have the protocol buffer message decoded. And instead of the field names, I have the field numbers here. ” So what you can do now, is craft a Create-TypeProvider gRPC request with this protobuf encoding, and changing the numbers to the different transport values you want to try.
And once created, you can then use the JSON API to list all TypeProviders and get the actual JSON name of it. And this way he figured out, the transport name has to be set to GSLB. Who would have guessed. And so now he created a TypeProvider, targeting the internal corporate issue tracker, specifically the REST API discovery endpoint, just to prove he can reach it.
29:47 - “So there is a method in deployment manager to list the types a specific type provider handles. And okay. This is what the issue tracker has. And this is live. The deployment manager right now, sent a request to the issuetracker API through GSLB and got it’s discovery document and processed it, and it’s showing us the results of what it found. this is the bug. I’m able to create typeproviders that talk to internal endpoints through GSLB, in this case an internal API.
But it could be any endpoint. ” Wow! Of course google has fixed this bug now. But this is such an amazing bug to me, because it involved so many small puzzle pieces. So many small tricks that have to come together. I learned a lot about Google internals from this, and hopefully it can help future bug hunters too. Ezequiel, congratulations again for winning. You really deserved it.
30:55 - But of course there were more cool submissions for the GCP Prize. So go head over to the blog about this year’s winners, and learn more about the other amazing bug hunters and their findings. .