Arbitrary File Read in HashiCorp go-getter Git URL Handling

This repository is a working PoC for CVE-2026-4660 affecting HashiCorp go-getter and downstream tools such as Terraform. It demonstrates an arbitrary local file read caused by passing an attacker-controlled Git `ref` beginning with `--pathspec-from-file=...`, which is later used in `git checkout` without a protective `--` separator. Git interprets the ref as an option, reads the specified local file line-by-line, and emits those lines in pathspec error messages, leaking file contents to terminal output and CI logs. Repository structure is small and purpose-built: `Dockerfile.gitserver` builds a malicious local HTTP git server hosting two bare repositories, including a legitimate-looking outer Terraform module (`terraform-aws-vpc.git`) whose nested module sources contain the malicious refs, and an inner module (`terraform-aws-vpc-internal.git`) used as the checkout target. `Dockerfile` builds the victim container with pinned vulnerable Terraform 1.14.8, fake AWS credentials, a fake SSH key, and the PoC runner script. `docker-compose.yml` orchestrates both containers and health-checks the git server. `poc.sh` is the main exploit entry point and demonstrates three stages: raw git checkout behavior, Terraform `init` triggering go-getter's `clone()` path, and a direct simulation of go-getter's `update()` path. Main exploit capabilities: (1) host attacker-controlled Git/Terraform modules over HTTP; (2) induce a victim to fetch a top-level module that silently references malicious nested modules; (3) leak arbitrary readable local files such as `/home/runner/.aws/credentials`, `/home/runner/.ssh/id_rsa`, and `/etc/passwd`; and (4) show behavioral differences between go-getter `clone()` and `update()` paths, specifically that `clone()` removes the destination on failure while `update()` leaves it behind. The exploit is operational rather than just illustrative because it fully automates environment setup and demonstrates real leakage through vulnerable tooling, but it uses hardcoded local infrastructure and sample files rather than a generalized payload framework.