Running policy engine may involve running untrusted, user-supplied code on user-supplied inputs. This document describes the security aspects that come into play.
This is less of a concern if you are only using policy-engine to test your own infrastructure against your own set of rules.
Even though security is important to us (given policy-engine originated at a security-focused company), it's not impossible to rule out bugs that would allow some sort of escalation.
Therefore, if you are using user-supplied Rego code or user-supplied inputs, we strongly recommend to run using policy-engine in an isolated environment, which typically means at least:
- No filesystem access
- No environment variable propagation
- No network access
- Strict limits on memory & time consumption
We use OPA's topdown Rego evaluator, written in Go. We have disabled the following functionality:
http.send(): can send arbitrary HTTP callsopa.runtime(): includes environment variables; this returns an empty object in policy-engine.
Passing resource information from .tf files involves evaluating HCL.
The hcl_interpreter README contains more
information on how this works: but what's relevant here is that we mix our
own evaluation logic with reusing bits of the hashicorp's HCL package.
Some functions in HCL allow reading information from the filesystem.
An classic example of this is something like:
resource "aws_instance" "my_instance" {
user_data = file("${path.module}/startup_script.sh")
}This would allow users to gain information about the server-side environment. To this purpose, we've modified the following functions to use the Afero filesystem library rather than the physical one.
- Filesystem functions:
abspathdirnamepathexpandbasenamefilefileexistsfilesetfilebase64templatefile
- Crypto functions:
filebase64sha256filebase64sha512filemd5filesha1filesha256filesha512