AWS Security Blog

How AWS uses automated reasoning to help you achieve security at scale

At AWS, we focus on achieving security at scale to diminish risks to your business. Fundamental to this approach is ensuring your policies are configured in a way that helps protect your data, and the Automated Reasoning Group (ARG), an advanced innovation team at AWS, is using automated reasoning to do it.

What is automated reasoning, you ask? It’s a method of formal verification that automatically generates and checks mathematical proofs which help to prove the correctness of systems; that is, fancy math that proves things are working as expected. If you want a deeper understanding of automated reasoning, check out this re:Invent session. While the applications of this methodology are vast, in this post I’ll explore one specific aspect: analyzing policies using an internal Amazon service named Zelkova.

What is Zelkova? How will it help me?

Zelkova uses automated reasoning to analyze policies and the future consequences of policies. This includes AWS Identity and Access Management (IAM) policies, Amazon Simple Storage Service (S3) policies, and other resource policies. These policies dictate who can (or can’t) do what to which resources. Because Zelkova uses automated reasoning, you no longer need to think about what questions you need to ask about your policies. Using fancy math, as mentioned above, Zelkova will automatically derive the questions and answers you need to be asking about your policies, improving confidence in your security configuration(s).

How does it work?

Zelkova translates policies into precise mathematical language and then uses automated reasoning tools to check properties of the policies. These tools include automated reasoners called Satisfiability Modulo Theories (SMT) solvers, which use a mix of numbers, strings, regular expressions, dates, and IP addresses to prove and disprove logical formulas. Zelkova has a deep understanding of the semantics of the IAM policy language and builds upon a solid mathematical foundation. While tools like the IAM Policy Simulator let you test individual requests, Zelkova is able to use mathematics to talk about all possible requests. Other techniques guess and check, but Zelkova knows.

You may have noticed, as an example, the new “Public / Not public” checks in S3. These are powered by Zelkova:

Figure 1: the "public/Not public" checks in S3

Figure 1: the “Public/Not public” checks in S3

S3 uses Zelkova to check each bucket policy and warns you if an unauthorized user is able to read or write to your bucket. When a bucket is flagged as “Public”, there are some public requests that are allowed to access the bucket. However, when a bucket is flagged as “Not public”, all public requests are denied. Zelkova is able to make such statements because it has a precise mathematical representation of IAM policies. In fact, it creates a formula for each policy and proves a theorem about that formula.

Consider the following S3 bucket policy statement where my goal is to disallow a certain principal from accessing the bucket:

    "Effect": "Allow",
    "NotPrincipal": { "AWS": "111122223333" },
    "Action": "*",
    "Resource": "arn:aws:s3:::test-bucket"

Unfortunately, this policy statement does not capture my intentions. Instead, it allows access for everybody in the world who is not the given principal. This means almost everybody now has access to my bucket, including anonymous unauthorized users. Fortunately, as soon as I attach this policy, S3 flags my bucket as “Public”—warning me that there’s something wrong with the policy I wrote. How did it know?

Zelkova translates this policy into a mathematical formula:

(Resource = “arn:aws:s3:::test-bucket”) ∧ (Principal ≠ 11112222333)

Here, ∧ is the mathematical symbol for “and” which is true only when both its left and right side are true. Resource and Principal are variables just like you would use x and y in algebra class. The above formula is true exactly when my policy allows a request. The precise meaning of my policy has now been defined in the universal language of mathematics. The next step is to decide if this policy formula allows public access, but this is a hard problem. Now Zelkova really goes to work.

A counterintuitive trick sometimes used by mathematicians is to make a problem harder in order to make finding a solution easier. That is, solving a more difficult problem can sometimes lead to a simpler solution. In this case, Zelkova solves the harder problem of comparing two policies against each other to decide which is more permissive. If P1 and P2 are policy formulas, then suppose formula P1 ⇒ P2 is true. This arrow symbol is an implication that means whenever P1 is true, P2 must also be true. So, whenever policy 1 accepts a request, policy 2 must also accept the request. Thus, policy 2 is at least as permissive as policy 1. Suppose also that the converse formula P2 ⇒ P1 is not true. That means there’s a request which makes P2 true and P1 false. This request is allowed by policy 2 and is denied by policy 1. Combining all these results, policy 1 is strictly less permissive than policy 2.

How does this solve the “Public / Not public” problem? Zelkova has a special policy that allows anonymous, unauthorized users to access an S3 resource. It compares your policy against this policy. If your policy is more permissive, then Zelkova says your policy allows public access. If you restrict access—for example, based on source VPC endpoint (aws:SourceVpce) or source IP address (aws:SourceIp)—then your policy is not more permissive than the special policy, and Zelkova says your policy does not allow public access.

For all this to work, Zelkova uses SMT solvers. Using mathematical language, these tools take a formula and either prove it is true for all possible values of the variables, or they return a counterexample that makes the formula false.

To understand SMT solvers better, you can play with them yourself. For example, if asked to prove x+y > xy, an SMT solver will quickly find a counterexample such as x=5,y=-1. To fix this, you could strengthen your formula to assume that y is positive:

(y > 0) ⇒ (x + y > xy)

The SMT solver will now respond that your formula is true for all values of the variables x and y. It does this using the rules of algebra and logic. This same idea carries over into theories like strings. You can ask the SMT solver to prove the formula length(append(a,b)) > length(a) where a and b are string variables. It will find a counterexample such as a=”hello” and b=”” where b is the empty string. This time, you could fix your formula by changing from greater-than to greater-than-or-equal-to:

length(append(a, b)) ≥ length(a)

The SMT solver will now respond that the formula is true for all values of the variables a and b. Here, the solver has combined reasoning about strings (length, append) with reasoning about numbers (greater-than-or-equal-to). SMT solvers are designed for exactly this sort of theory composition.

What about my original policy? Once I see that my bucket is public, I can fix my policy using an explicit “Deny”:

    "Effect": "Deny"
    "Principal": { "AWS": "111122223333" },
    "Action": "*",
    "Resource": "arn:aws:s3:::test-bucket"

With this policy statement attached, S3 correctly reports my bucket as “Not public”. Zelkova has translated this policy into a mathematical formula, compared it against a special policy, and proved that my policy is less permissive. Fancy math has proved that things are working (or in this case, not working) as expected.

Where else is Zelkova being used?

In addition to S3, several AWS services are using Zelkova:

We have also engaged with a number of enterprise and regulated customers who have adopted Zelkova for their use cases:

“Bridgewater, like many other security-conscious AWS customers, needs to quickly reason about the security posture of our AWS infrastructure, and an integral part of that posture is IAM policies. These govern permissions on everything from individual users, to S3 buckets, KMS keys, and even VPC endpoints, among many others. Bridgewater uses Zelkova to verify and provide assurances that our policies do not allow data exfiltration, misconfigurations, and many other malicious and accidental undesirable behaviors. Zelkova allows our security experts to encode their understanding once and then mechanically apply it to any relevant policies, avoiding error-prone and slow human reviews, while at the same time providing us high confidence in the correctness and security of our IAM policies.”
Dan Peebles, Lead Cloud Security Architect at Bridgewater Associates


AWS services such as S3 use Zelkova to precisely represent policies and prove that they are secure—improving confidence in your security configurations. Zelkova can make broad statements about all resource requests because it’s based on mathematics and proofs instead of heuristics, pattern matching, or simulation. The ubiquity of policies in AWS means that the value of Zelkova and its benefits will continue to grow as it serves to protect more customers every day.

Want more AWS Security news? Follow us on Twitter.