Hacker News Reader

Comment by egorfine | Hacker News Reader

I have been discussing IAM with a couple of AWS engineers from that team a few years ago. I have tried to convince them to implement something actually sane non-corporate developers, because IAM is complicated way beyond a point of being usable. I have seen so many people just slam the root credentials into config in order to never touch IAM. And yes, me included, and yes, it's a shameful practice, don't do that. Take a day off, go on a hike, watch the sunset, meditate and embrace the inevitable: yes IAM is incredibly bad and there is no way around.

(I was heavily using AWS since inception, way before IAM was invented).

My speech fell on deaf ears, because IAM is clear as sky to its inventors and they fail to see what's the problem in it and corporate customers are happy to employ one or two people solely dedicated to learning and managing IAM.

Anyways. There is no way around IAM. And just like with systemd you better embrace it and learn it and it turns not that terrible.

View on HN

> I have seen so many people just slam the root credentials into config in order to never touch IAM

Anecdotally, at a startup I worked for, our devops and security engineers pulled me into a call once because they needed my help finding where the root credentials were being used in our code. I logged into our dev, staging, and production servers, but they were nowhere to be seen in any place you'd expect AWS creds (like ~/.aws).

Then it hit me. I pulled up our git repo and pulled up a function that handled file upload to s3 for something incredibly unimportant. Right there, in plain text, checked into git. The git blame read something like "4 years ago".

Creds weren't even needed since the AWS SDK does whatever magic to know the creds based on the EC2 instances role. I immediately put up PR just removing the creds as the empty constructor for our SDK used the EC2 role (this pattern was used in multiple other parts of the code base).

The CEO (the author of the code) didn't want to merge it in case it stopped working, despite me demoing in our staging environment and the feature using this code was only used like twice a week.

Thankfully, we convinced him, got it removed from the codebase, and either cycled or deleted the root creds, I don't recall how that worked at the moment.

I wasn't familiar with root creds at the time either, but damn is it scary to think that anyone of the engineers had access to root creds for a company of the size that we were.

And the CEO wrote it like that originally because he didn't want to mess with IAM. I can see why he did it, but man was that a bad idea.

Not a good sign...

> And the CEO wrote it like that originally because he didn't want to mess with IAM. I can see why he did it, but man was that a bad idea.

This is a very common case. It's okay in the very early development (rarely we start projects with a definition of IAM policies), but then it gets shoved into backlog forever.

There’s a pattern I see (and indeed have even perpetuated once or twice) of “AWSguide based development” where someone copy pasted the default policy/roles directly from a guide into AWS CLI commands and just forgot about them after. Not as bad as the root credentials being used but certainly doesn’t help maintain a sane and understandable access system.

I know it’s a lot of extra maintenance effort and probably more difficult but AWS could go a long way towards making this more sane and transparent by giving gitops examples alongside their CLI ones. Even someone just copy pasting terraform or cloud formation is going to be way more visible than AWS CLI

I'm not sure how it is ever ok to put your _root_ credentials into version control? I am not even comfortably doing that with my personal private repositories that I share with nobody.

Oops, I replied to an out-of-context comment. I mean it's okay to use a wildcard policy in the early stage of a project.

Of course, committing credentials into a repo is absolutely never ever ok.

I disagree. IAM does what it’s supposed too when you need to run an organization with very fine gained controls.

People being able to see and interact with certain services but not everything related to those services. Same with services since it’s the same system.

To me the problem isn’t IAM itself but the nested API calls. Which also itself isn’t a problem but we need a better UI to diagnose when something goes wrong.

Cloudtrail isn’t really suitable for this task. A “simple” UI of looking at a role and seeing a “failed calls” would go a long way.

There is a pretty simple way to query cloudtrail logs with Athena and its a huge improvement over the normal UI.

I probably should have been more clear on that, it is very possible with Cloudtrail and Athena and I find myself doing that pretty regularly.

But there are also times that it is inconsistent at best especially when trying to look at some nested permission problem. More than a few times I have had to get on with AWS support because the actual error just was not in Cloudtrail anywhere. Or it is related to some service that doesnt log to Cloudtrail like s3 access.

Which kinda more my point was, it isnt IAM itself that is the problem.

When "run a SQL query against an implicit file database" is a huge improvement for anything but the edgiest of cases (and I believe you that it is), that's a scathing indictment of the normal UI.

What I don't understand is how the AWS log inspection tools are still as bad as they are. Even if it's just to prepare public-facing material, AWS clearly dogfoods them a little bit, so surely there would be glory and accolades to be won by implementing search that was half-assed (instead of quarter-assed)? Or is the AWS culture so broken that it net punishes core improvements? Come to think of it, that would explain a lot.

I feel like IAM is not actually that complicated? Most services have permissions that are in direct correspondence with the API calls you want to make. Of course a few of them decided to do mad shit (looking at you API Gateway), and a few others require you to have nearly all permissions to do some of their most common API calls (S3 needs like 15 permissions to set all possible properies on an S3 object).

It’d be nice if AWS could fix the madness and standardize on one List/Update/Create/Delete call per resource, since most services are already like that.

Don't they have a table that states api call -> IAM permissions necessary?

That would go a long way into solving the problem

I find IAM to be a great product. I cannot imagine how we could make it different, without removing some of its capabilities: after all, this is nothing but a list of <who, what, on which> permissions

> I cannot imagine how we could make it different

This seems almost bait.

Problem is, once you dive deep in the rabbit hole, you will inevitably build another IAM.

The art here is knowing when to stop building, not what to build.

This is a great way to avoid voicing constructive criticism about IAM that doesn’t end up being “hard-code it for my use case at the expensive of making it difficult for a significant portion of other customers”.

Well, for a start, you could have a recorder that actually shows all the relevant permissions for a particular API call so users can work out what they should set permissions on (including interactions with SCPs and anything else that could affect a decision).

You can easily get the first part of via the sdk https://cloudonaut.io/record-aws-api-calls-to-improve-iam-po....

You can actually set this up in your account. It’ll record all the permission your roles actually use (over a customizable period of time), so you can remove all others

I tend to agree with you that the idea of a list of "who, what, which" is straightforward and probably is the right way of doing things. I agree that permissions is inherently a hard problem.

I disagree that IAM specifically is a great product. It's great if you learn it, but learning permissions system never adds anything to the bottom line. It's an evergrowing technical debt at best. People interested in learning permissions system are typically people interested in developing one. Users are never happy with permissions systems because more often than not it's a nuisance.

In my view IAM is an incredibly overengineered list of who, what, which, suitable for corporate granularity but then too complex to be used by a startup or a project.

Yes

You are diving into another topic: "fine-grained permissions"

In my opinion, fine-grained permissions is a very good idea for application-attached role: your lambda can only do "this", your ec2 instance can only do "that". For humans, they are a mistake and shall be avoided.

I'd love for fine-grained permissions to be optional. Not an insurmountable barrier to the actual service.

Like let me start the development first. I'll get to security later. I came here for object storage, don't force me to learn an incredibly complex auxiliary service.

But they are, that's kind of the whole point. You start with {"actions": ["s3:"], "resources": [""]} and graduate from there when ready.

This could be a perfect 101 on IAM which I would strongly prefer never attending.

How else do you expect to manage permissions? Is there a better dsl design you have in mind?

What is better? I am pretty sure IAM is the best there is, no questions about that.

Better for small teams? Sure: a login/password system with maybe optional restrictions on services and maybe "readonly" access and maybe with disabled aws console.

Basically, something that was in AWS before the IAM madness came in.

But you can do that with IAM, and quite easily too

Create IAM user with access/secret keys are easy

Assign the built-in ReadOnly policy, easy too, or any other built-in "generic" policy

You are missing my point.

Of course, it is possible with IAM because IAM is a top-tier ACL system.

The point: I don't want to learn it.

If my account has only got one thing in it, I don’t expect to manage permissions at all.

“Everything has root” is fine when you are starting a new thing - what’s needed is tooling to start there and grow into fine grained permissions when you add a second thing.

> Like let me start the development first. I'll get to security later.

And that's why the world is full of crappy, insecure software. Security is ever an afterthought.

Wholeheartedly agree with you. Not only it's an afterthought, it's a sinkhole for resources with zero apparent ROI. It's basically lives in backlog rent-free.

They basically are optional. You can use wildcards in the action. (Nobody will go back and tighten them up though.)

They aren't. To have wildcards I still have to go to IAM and learn how to create a wildcard policy.

But you don't have to list each and every permission, which was my point. This is far easier than investigating what your app does, mapping it to permissions, missing some of them, and iterating...

The true purpose of IAM is for AWS control-plane and data-plane elements to interact autonomously with resources that are in your account without violating the dialectic around security “of the cloud” vs “in the cloud”.

And this is also why you need a PhD to fully comprehend it

No. You’re in denial about the importance of authorisation, don’t want to give it the time it deserves, and are shooting the messenger in placing the blame on IAM to the degree that you are. People see AAA as a pain in the ass because it’s not the problem they thought they’d need to solve. That doesn’t mean that it’s important, or that the complexity is avoidable.

For the record, I’ve never been in a development team of more than 4 people, and my teams have always been responsible for their own ops. No ‘corporate dev’ here.

> People see AAA as a pain in the ass because it’s not the problem they thought they’d need to solve.

True.

> That doesn’t mean that it’s important,

Not necessary true.

> or that the complexity is avoidable.

Untrue. There is no need for IAM for the vast majority of 4 people teams and there could be a simpler solution for those, like it was before IAM. And it can actually be built on top of IAM, just spare people from diving deep into it.

> suitable for corporate granularity but then too complex to be used by a startup or a project

To some extend. It’s perfectly possible to make a role that allows S3 access to any action on everything, but still restricts the service from operating on any of the other 400 AWS services (unlike root credentials)

I don’t think the language is bad, but maybe improve the policy simulator, make it easier to understand what policies are require and also make it easier to find and reference different arns.

I wish it is a list of <who, what, which>. Right now you can't even get a comprehensive list of people who can access this S3 object. Add assume roles, conditional access, cross account, per-service object level policy into the mix I don't think it's even actually possible for AWS to implement that.

I found GCP's access control to be much simpler but more limited than AWS. They're adding complexity though with conditional access. In GCP the "who" in most case is an email (Google Account, Google Groups, service account)

You are right : "much simpler but more limited"

And because you can limit yourself, you can effectively simplify your way of consuming IAM

"He who can do more can do less"

The way it evolved makes it complicated. That you have normal policy, resource attached policy, SCP, Assumed roles, etc, etc.

That spreads out the actual end resulting policy into a bunch of disparate corners that's hard to unravel.

There's tools to see "can this specific user/role/service get to this thing". But because of the above, there's no one place to see "who can/cannot get to this thing".

I'm not sure how that is complicated. There are different systems with different functions for different purposes, that will not go away unless you remove the systems or the purposes. And that's not useful.

Okay yes, I figure out what I need to. But it seems clear that conflicting concepts got bolted on after the initial design. Which means it's more complicated than it needs to be.

Which ones are bolted on? Besides the S3 ACLs (which are more like a precursor) and the weird conditional language it's all pretty consistent and obvious like any ACL.

Things like principals, resources, trusts and control policies are all distinct systems with different goals and purposes. Maybe I'm missing some different AWS IAM policy that has that bolt-on flavour?

You can't meld the rules together to get a cohesive view of, for example, "who can access this S3 bucket?". Because, for example, an SCP can override a bucket policy. And policies can be other places too.

I can get "can this specific role/user access this specific bucket", just not the other direction.

It's not because the language isn't consistent, it's because the language isn't backed by a single cohesive RBAC type setup.

I'm not sure how that makes it a bolt-on or conflicting or complicated. It is indeed not a simple central spreadsheet of 'who can do X on Y', but that would not be feasible at AWS scale.

Microsoft tries that with Entra and it's a pretty miserable experience. Google has CEL for the conditionals which is neat, but it gets rather close to software engineering rather than IAM configuration at that point. The somewhat hierarchical nature they use is also not as great as it seems on first glance.

Yeah I'm not saying I can't use it. I'm just trying to explain why people might look at it and feel like it's not unified.

Permission boundaries

> There are different systems with different functions for different purposes,

Which are interconnected, right? If yes, then that's what the word "complicated" means: made of multiple, non-trivially interacting parts.

You can choose to interconnect them but that is neither required nor default. It depends on what you want to build.

You can have a simple policy on a resource that just allows something, or you can make a complicated bi-directional set of policies that refer to specific principals, resources and actions to constrain it more. That will never go away if that is what you need to build. And it is still optional because if you don't want to build that, then you just end up with your single, not-inter-connected policy.

Imagine, you drop IAM and revive the good old user/password system.

You have no way to see "who can/cannot get to this thing". Yet who could argue that user/password are "complicated" ?

NB: using IAM access/secret key has indeed the same issue

I'm not arguing for user/password.

AWS is good at building abstractions on top of their already existing features, e.g. AWS App Runner. Maybe there could be an abstraction on top of IAM that does not allow for all the fine-grained permissions settings and options that IAM has, but has much simpler predefined roles which are very well suited for a small startup.

For a lot of situations they already have this. They have managed (AWS Created and Managed) policies that if you want a basic level of security you can use.

A lot of services will create policies for you, for example you can go to RDS and click setup connection to lambda, or ec2 and it will create the policy.

A lot of things will give you the policy to copy and paste.

Another UI to further abstract IAM would likely just complicate things, and then make it harder later if/when you need to leave that abstraction.

The main problem is it's a huge time suck. When writing IAC like Terraform, you'll probably spend more time configuring IAM roles and permissions than you do the actual infrastructure. Also, in my experience, nobody reviewing your code will check if it's correct, least privilege, etc. And when you get it wrong (too restrictive), the error messages can be difficult to diagnose, depending on the service. So most people will err on the side of more permissions than needed.

I have hard this from many developers. Are there any good solutions or best practices I should look into?

For the security sanity checking of terraform infra code, take a look at Checkov.

Don't use Terraform? If you want to create individual resources one by one, go for it. Use something like CDK which abstracts a lot of the boilerplate away.

I wound up with a job at a TF-heavy company. I personally despise Terraform. The language is anemic and there's too much copy-paste code.

…and compiles into CloudFormation, which is of the devil.

I also have issues with infrastructure being declared imperatively. I’ve seen countless bugs and incidents caused by CDK surprises.

You get precisely what you ask for with a declarative language.

It’s a little more complicated than that. I’ve run into many instances where a particular Action involves many different Resources.

If you’re trying to do fine grained permissions you have to know every resource that might end up in the request context and then have an entry in a Resource section of a policy for that Resource type. For instance just creating an EC2 (RunInstances) might involve Subnet, Security Group, EBS, image, volume, snapshot, and a few other resource types.

This gets even more complicated when you’re using Condition operators because the conditions are checked against every resource in the request context but each resource type has different valid conditions (eg some resources have tags and some don’t) so you have to split up the policy into numerous statements specific to each resource type.

All of that would be fine if there was any way to actually reliably get the context of a request so you can see exactly what was denied. Sometimes you can get the context via the encoded authorization message in a failed API call but this is usually truncated in CloudTrail or if you’re using IaC tools like Cloudformation.

They’ve done a lot in the past few years to explain why a particular call was denied but none of it beats seeing the context and seeing exactly which part of the policy did or didn’t apply.

I think this gets at my ick with IAM. If AWS knows it needs “s3:ListBuckets” for a call to function, why doesn’t giving permissions on that call just imply that I gave it the “s3:ListBuckets” permission too?

For enterprises that genuinely want the finer grained control, let them express that they want to opt out of that implicitness in the policy document.

I think it gets hard when an emergent chain of complex trust relationships need to be built and understood. Things like IAM identity center, workload identity, IRSA on EKS, service principals vs roles for accessing other services from a service, resource policies vs principle-level policies and when to use each. Not necessarily intuitive all the time. I don't think it's THAT hard and I understand why some of these things were built this way, but it's a huge complicated ecosystem of services and I understand why it can get confusing to some. Gotta be disciplined about it.

It's not THAT hard if that's the thing you specialize about.

Problem is, that kind of things are the backend of the backend. An accessory of an accessory. No business cares about it. Ever. And yes, running wildcard policies can hit you hard if not attended properly; my point is I don't want to learn a complex ACL system made for enterprise for my small startup that is actually gonna be fine with wildcard policy basically forever.

“Record mode”, where for 24h all permissions are automatically granted on first use, would be an obvious improvement.

The ability to measure what permissions you need and record them automatically would improve the onboarding experience immeasurably.

In my experience, it's fine. I just wish it supported more condition keys across all services. Nothing sucks more than to write a policy with a explicit condition only to bash your head trying to figure out why and then finally realize that restricting for a specific tag for a specific resource+action is not supported.

It's funny, because AWS devs are also a bit guilty of slamming admin credentials and calling it a day.

I was testing out sagemaker studio. I just did the quick setup wizard and the default managed execution role was insanely permissive. I believe read/write to all of the accounts S3 buckets and broad List* for account resources. There are multiple parts of the documentation that also recommended you use this role. It seemed especially wild for a product with so many ways to access. we have good account hygiene but still

> the default managed execution role was insanely permissive.

I find this is often the case.

Cedar, the new language by AWS that's currently used by Verified Access, is looking good; https://www.cedarpolicy.com

Their mere description on that page makes me shiver.

this is application side, correct?

I think IAM would be a lot easier to interact with from a dev perspective if the error messages were more clear. I suspect that they are sometimes intentionally vague (or misleading) as a kind of security-through-obscurity. For example, S3 returns “file does not exist” errors when the real problem is that the caller doesn’t have permission to access the bucket/prefix.

I'm not sure how IAM is any more difficult or harder to comprehend than any alternative. I've been using AWS since 2010 and frankly IAM has never been an issue. Can you give some examples?

In any case, IAM and its essential features are not some optional thing; it's actually the prerequisite for AWS existing in its current form. IAM is the most universal AWS service - every other AWS service must use it. It's a loosely coupled symbolic policy computation engine for federated identity and service permissions. The "federated" part is doing A LOT of heavy lifting here: without federation, it's impossible to decouple teams and achieve the organizational velocity that AWS has.

Other cloud providers have ended up re-inventing AWS IAM, sometimes poorly. Although I have to give GCP credit for recently greatly improving their IAM console and permissions error usability.

IAM is conceptually straightforward, but it's combinatorial difficult with all the different ways that AWS services interact with it. Support for IAM is also inconsistent across services. I'll give a recent example.

I have EC2 Mac instances in an autoscaling group that are running untrusted code. I've "sandboxed" that code by running it as its own user and blocking that user from reaching the IMDS endpoint with a packet filter on the instance so that the untrusted code, in theory, cannot get instance policy credentials. Nonetheless, I'd like to limit the blast radius should the code somehow escape that sandbox. So I want the instance policy as restricted as possible.

One of the things the instances need to do is set tags on themselves. For this, you can use an IAM condition on the ec2 CreateTags action to say "this instance can call CreateTags, but the instance ID has to be the instance itself". This prevents an instance from setting tags on a different instance in the same autoscaling group.

The instances also need to make a few autoscaling API calls, to complete lifecycle actions (CompleteLifecycleAction), and to enable and disable scaling protection (SetInstanceProtection). Here again, I'd like each instance to only be able to make these calls against itself. Unfortunately, there is no condition you can use with the autoscaling API to similarly limit the calls as you can with the ec2 API. The same condition that works with the ec2 API doesn't do anything with the autoscaling API.

Not helping matters, the documentation is very spread about. Just in the use case I describe above, you need to read at least all of these:

https://docs.aws.amazon.com/service-authorization/latest/ref...

https://docs.aws.amazon.com/autoscaling/ec2/userguide/ec2-au...

https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/iam-poli...

https://docs.aws.amazon.com/autoscaling/ec2/userguide/contro...

BTW, there's no reason I can see that the condition couldn't work in both cases, but apparently this is up to the API/service to implement. The condition for CreateTags is:

  Condition:
    StringLike:
      aws:userid: "*:${ec2:InstanceID}"

Documented (sorta) across four(!) pages:

https://docs.aws.amazon.com/IAM/latest/UserGuide/reference_p...

https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/iam-poli...

https://docs.aws.amazon.com/service-authorization/latest/ref...

Reply on news.ycombinator.com