CVE-2026-32805 — webserver

Summary

The sanitizeArchivePath function in webserver/api/v1/decoder.go (lines 80-88) is vulnerable to a path traversal bypass due to a missing trailing path separator in the strings.HasPrefix check. A crafted tar archive can write files outside the intended destination directory.

Vulnerable Code

File: webserver/api/v1/decoder.go, lines 80-88

func sanitizeArchivePath(d, t string) (v string, err error) {
	v = filepath.Join(d, t)
	if strings.HasPrefix(v, filepath.Clean(d)) {
		return v, nil
	}
	return "", &ErrPathTainted{
		Path: t,
	}
}

The function is called at line 48 inside [*Decompressor].Unzip, which is invoked by Decode (line 80) during execution of the webserver CLI (command download).

Root Cause

strings.HasPrefix(v, filepath.Clean(d)) does not append a trailing / to the directory prefix, causing a directory name prefix collision. If the destination is /home/user/extract-output and a tar entry is named ../extract-outputevil/pwned, the joined path /home/user/extract-outputevil/pwned passes the prefix check — it starts with /home/user/extract-output — even though it is entirely outside the intended directory.

Steps to Reproduce

1. Deploy Romeo. A measured app writes its coverage data.

2. Place the PoC zip on the PVC. Any pod with write access to the ReadWriteMany PVC (or the webserver itself) copies a poc-path-traversal.tar into the coverdir mount path. The archive contains legitimate coverage files alongside two crafted entries with path-traversal names.

3. Run the webserver CLI against the running webserver:

   webserver download \
     --server http://localhost:8080 \
     --directory /home/user/extract-output
   

4. Observe the bypass. unzip processes the zip stream. For the malicious entries:

   // entry name: ../extract-outputevil/poc-proof.txt
   filepath.Join("/home/user/extract-output", "../extract-outputevil/poc-proof.txt")
     => "/home/user/extract-outputevil/poc-proof.txt"
   
   strings.HasPrefix("/home/user/extract-outputevil/poc-proof.txt",
                     "/home/user/extract-output")
     => true   // BUG: prefix collision; file lands OUTSIDE target dir
   

   Both malicious entries are written outside /home/user/extract-output/. The legitimate coverage files land correctly inside it.

Impact

Successful exploitation gives an attacker arbitrary file write on the machine running the webserver CLI. Real-world primitives include:

• Overwriting ~/.bashrc / ~/.zshrc / ~/.profile for RCE on next shell login
• Appending to ~/.ssh/authorized_keys for persistent SSH backdoor
• Dropping a malicious entry into ~/.kube/config to hijack cluster access
• Writing crontab entries for persistent scheduled execution

The attack surface is widened by the default ReadWriteMany PVC access mode, which means any pod in the cluster with the PVC mounted can inject the payload — not just the Romeo webserver itself.