Azure's Weakest Link? How API Connections Spill Secrets
binarysecurity.no
binarysecurity.no
Performance, "I don't like the portal", service and capacity availability, and such complaints are somewhat subjective or fixable but I deeply believe Microsoft is the most insecure of the cloud giants on a measurable level.
Anyone that is serious about security should just avoid Microsoft, this has honestly been the case since the early '00s at the least.
UPD: note to self - this seems like a good resource https://www.cloudvulndb.org/results
• Storm-0558 Breach (2023): Chinese hackers exploited a leaked signing key from a crash dump to access U.S. government emails, affecting 60,000+ State Department communications
• Azure OpenAI Service Exploitation (2024): Hackers bypassed AI guardrails using stolen credentials to generate illicit content, leading to Microsoft lawsuits against developers in Iran, UK, and Vietnam
• CVE-2025-21415 (CVSS 9.9): Spoofing vulnerability in Azure AI Face Service allowed authentication bypass and privilege escalation
• CVE-2023-36052: Azure CLI logging flaw exposed plaintext credentials in CI/CD pipelines, risking sensitive data leakage
• Azurescape (2022): Container escape vulnerability enabled cross-tenant access in Azure Container Instances, discovered by Palo Alto Networks
• ChaosDB (2022): Wiz researchers exploited CosmosDB’s Jupyter Notebook integration to access thousands of customer databases including Fortune 500 companies
• Executive Account Takeover Campaign (2024): Phishing campaign compromised 500+ executive accounts via Azure collaboration tools with MFA manipulation
If your company or workplace is considering migrating from cloud to on-prem or from one cloud to another, I do this professionally btw, feel free to reach out at this temporary email and we can chat: pale.pearl2178 at fastmail.com (to prevent my real email being scraped from HN).
For me it's just a distant dream now, but I bet business will be booming for you in the coming years, especially if you're located in Europe ;)
If these issues remain unfixed after being disclosed, or a pattern of fixes that took much longer than you feel they should have, that's valuable ammunition as it shows the organization isn't responsive to security issues.
Take something like a container escape vulnerability.
We could have Vendor A where they're just running containerd on a bunch of hosts on a single network segment and throwing everyone's containers at it so a container escape vulnerability essentially gets you access to everything any of their customers are running.
Where-as Vendor B segments running containers into VMs, so a container escape vulnerability means you can only access your own data. Not great because if one container is compromised that gives them a path into the rest of your workloads, but at least I know they're maintaining a pretty solid wall between tenants.
Then there's Vendor C that actually runs containers using some micro-VM framework so each container is running fully isolated by a hypervisor with a fully separate emulated network stack, etc so the escape really gets them no more access than they had inside the container.
A pattern of issues like Vendor A is, well, a pattern. A series of issues that show their systems are fundamentally not designed for proper isolation between tenants and are lacking defense-in-depth measures to mitigate the fallout of the inevitable security issues is a very good reason to write off Vendor A regardless of how quickly they respond to the issues.
I'm not going to go back and review all the Azure issues, but my recollection from the few writeups I've read definitely paint a picture of a lot more "Vendor A" type issues than I'd be comfortable with.
I’ve been there, done that, and was amazed how the security aspects only rapidly escalated to many millions of dollars and an ongoing cost also in the million or two range!
Think of this like a CEO: they’re less worried about Chinese hackers and more worried about about insider attacks. They’re much more common and do way more financial damage.
The cloud automatically provides separation of roles because an entirely different vendor is in charge of the lower layers, such as networking and storage.
Do you have any idea how hard it is to prevent a smart sysadmin from simply copying all data to a USB drive and walking out of the building with it?
That’s much harder when everything is on a managed hosting platform and no single person can access all accounts / subscriptions.
No, this thread is about Azure in particular having a bad security posture, not the cloud in general.
With Office 365, for example, they had at least 4 government clouds, some of which used shared infrastructure with Azure commercial, but had different data residency or employee requirements. They have thousands of employees monitored by all of the states as a condition of working on those clouds, for example.
Technical controls are similar, but the weak point are things that can cross cloud boundaries. One of the Chinese breaches of US government systems were caused by a PKI vulnerability that allowed the attacker to pivot from a dev environment to a federal cloud instance.
The deep integration with AD (now Entra) was the strongest selling point for Azure, but it’s also by far the biggest issue with the platform IMO.
There’s also just no consistency in the platform - the CLI for instance has totally different flags and names depending on which sub command you’re using. It’s like this everywhere in Azure.
For all of AWS's faults, one of the reason I really like them is how consistent everything is. There were so many instances where I could correctly guess the right command for the AWS CLI based on how other services worked, I could never do that with GCP or Azure.
I would love to read an article about how AWS ensures this kind of consistency. Given how Azure and GCP both messed this up, it's clearly not a trivial problem (even though it may seem like one)
* focusing on resources and operations on resources
* using consistent and expected naming schemes, pluralization, etc.
it also helps that the sdks and clis are very raw wrappers around this, such that if you know what it looks like in the sdk then it will look similar in the cli.
So the doco and the UI ends up littered with things like:
PrincipalId (ClientId)
For single tenants this might seem confusing, because you have both for a single app.
But if you were to have multi-tenants apps, each tenant would have their own Enterprise App instance, all referencing the same App Registration.
appId is for App Registrations.
objectId is for Enterprise Application Registrations.
clientId will be same as appId. It is used in the context of authentication, where it is the id of the object as client.
“EnterpriseAppId” and “AppRegistationId” would make sense.
ObjectId is meaningless nonsense. Everything is an object! Everything has an Id! This tells you nothing specific.
ClientId again is the id of the client, which does not have to be an app registration specifically.
I do agree it can be very confusing
You should make a YouTube channel in the style of 3b1b.
They also have a lot of different resources, such as Graph API, Entra ID.
Manage identities are simpler, since they are Azure constructions, so they work more or less like a IAM role. But then you try to use them with Entra ID APIs and things fall apart.
Microsoft's corporate culture evolved during some two decades under a very different threat model & security posture. A lot of the platform's foundations originate from that era, and although they've made significant strides you still see some of their other products struggle in this aspect as well (eg. good security primitives are there but they're sloppy in their default configuration and intuitability). Compare it to a platform that spent its youth growing up in a more actively hostile environment (like Bitcoin).
Well, the perimeter is not a gate but a cattle guard, and I am not surprised to see some wolves eating a secret and a cow swaggering into the road.
Azure service APIs have always conflated the principles of “reachability from the public internet” and “anonymous access” into a single concept called “Public Access” which, for Azure KV, has 6 different public/private configuration combinations!
This vulnerability report did not include the Key Vault Networking settings for “Public network access”, so more testing (but not much more) is needed to see if the proxy side door can circumvent a resource ACL or private endpoint or both.
Yeah, no joke. Considering how well protected Azure Key Vaults typically are, and what's in them (secrets, certificates etc) this is huge way to compromise a lot of other things. It's finding the keys to the doors.
:-)
"Microsoft confirms partial loss of security log data on multiple platforms" - https://www.cybersecuritydive.com/news/microsoft-loss-securi...
"Microsoft called out for ‘blatantly negligent’ cybersecurity practices" - https://www.theverge.com/2023/8/3/23819237/microsoft-azure-b...
So from that I can imply there was no payment.
All I got was an email: that is interesting bye bye. But it was fixed in the next patch or the next after I think.
So I didn't care to report the two bigger problems I found with Azure Information Protection [1][2] I thought about reporting them but decided against it.
And I will continue to tell people that I don't care to do free work for MS when they won't even give me a t-shirt, a mug or even acknowledge it.
Maybe if one is a security researcher it can be worth it but if you just find something interesting you'll probably be better rewarded by reddit or HN, yes, the upvotes are worthless but less so than a dismissive email.
[1] one in the downloadable AIP tooling where you can easily smuggle clear text information with rock solid plausible deniability - I found it by accident after having implemented a part of a pipeline in the most obvious way I could think of.
[2]: the second had to do with how one can configure SharePoint to automatically protect files with AIP on download, the only problem being if you logged in using another login sequence (sorry for the lack of details, this was before the pandemic and it was just a small part of what I was working on at the time) SharePoint would conveniently forget all about it despite all efforts by me, the security admin at the company and the expert that Microsoft sent to fix it.
Ha ... ha ... ha ... ha ... did they give you the run around for several months until you dropped the issue? It's actually pretty astounding that they don't get sued for this practice. If a company is paying for support and are given illiterate noobs then that is breach of contract I would think. I would never recommend entering a contract with MSFT, they produce trash products they can't support and are more invested in their Legal team than actual product.
Think of the cost of opportunity in having smart, capable, experienced staff doing support or backports instead of actual dev work. (Especially backports, which when they were done frequently they were done precisely because customers are risk-averse, so a great deal more review and testing (with a much larger test matrix) was required for backports, with attendant huge increase in cost.) That cost is enormous. But of course they do need to provide some support, and at some point some really good support for the really serious bugs, and the vendor will in time do it, but first the customer demand and pressure has to build.
I feel like I know more about M365 than anyone I talk to at MS. That's bad.
Azure basic support is pretty bad. I am even surprised at how bad some of the higher tiers of support are. I have no frame of reference for other cloud providers but anecdotally they are not so great either (some CSP's I am told are worse).
All in all - wrt documentation - software docs have gone downhill across the industry since the 90's. And back then people were still loudly grumbling about how bad the docs were!
Managers on the support side or your teams?
I had a friend who worked for an Cable ISP decades ago in the UK. Management of support management got outsourced to another company, who set aggressive targets for call length. Not average call length, but call length of any call they received. Any call that went over the target was a mark against the support person, and if you got more than a few marks you got a dressing down by the supervisor, a few more after that would get you a written warning, and then a few more would see you fired.
It started out at 15 minutes, and that was okayish. It took about 6 minutes to reboot a cable modem and have it come on-line, and that was done with almost every single support case, and fixed at least half of them.
Then they cut it down to 10 minutes. That was squeezing it a bit. 4 minutes at the most to do all introductions, hear the problem, wait for modem reboot and test things were resolved.
Then they cut it down to 5 minutes. The support folks had literally no choice but to just randomly hang up on people as soon as they got close to 5 minutes, or ask them to do a reboot of the modem and phone back. "Oh, I'm sorry, we must have been randomly disconnected"
Honestly I’m surprised they even acknowledged that as a bug, given there are many ways to get a whole lot more info than what you demonstrated, for instance the builtin “eye” button that is purpose built to reveal the full password to anyone with physical access to the machine wishing to see it.
This wasn't such a case.
That said, I didn't expect it to get rich, it was just that the experience didn't give me anything back for the effort I put in.
I'm glad they fixed it, but this doesn't seem too scary??
If user U can gain access to keyvault K via this exploit, it is scary.
[Vendors/Contingent staff will often be granted read-level access to a subscription under the assumption that they won't have access to secrets, for example.]
(I'm open to the possibility that I'm misunderstanding the exploit)
The secret typically is a user account (OAuth token), but it could also be an App Id/Secret.
It seems to me the KeyVault secret leak originated when KeyVault K owners gave secret reader permissions to the API Connection. (And I will note that granting permissions in Azure requires Owner role-which way more privileged than the Reader role mentioned in this article.)
[edit - article used Reader role, not Contributor role]
> The Inherent Insecurity of API Connections
> anyone with Reader permissions on the connection is allowed to arbitrarily call any endpoint on the connection
My concern here is that a lot of people have the takeaway that APIs shouldn't be exposed because they create security risks. But that's not true. The API exposure isn't the risk, it is the access control. If you don't have proper access control then it really isn't going to matter if you have an API or not. But then again, we have a long history of not taking fairly basic security seriously and with decades of computing and seeing the results, I really can't figure out why. Sure, security is expensive, but bad security is far more expensive. I guess maybe the issue is I'm not much of a gambler.
> it is common to not mark input (and output) as sensitive.
1) Fail open: default setting is that things are not marked as sensitive and an active decision has to be taken to mark sensitive
2) Fail closed: by default things are marked sensitive and action needs to be taken to mark it as non-sensitive
> What I would not expect is that anyone with Reader permissions on the connection is allowed to arbitrarily call any endpoint on the connection:
> all Readers on that subscription can call all GET requests defined on the connection.
What am I missing here?