GHSA-v856-2rf8-9f28

Summary

A crafted DICOMDIR can set ReferencedFileID to a path outside the File-set root. pydicom resolves the path only to confirm that it exists, but does not verify that the resolved path remains under the File-set root. Subsequent public FileSet operations such as copy(), write(), and remove()+write(use_existing=True) use that unchecked path in file I/O operations. This allows arbitrary file read/copy and, in some flows, move/delete outside the File-set root.

Details

Verified on pydicom 3.1.0.dev0.

Relevant logic is in src/pydicom/fileset.py:

• RecordNode._file_id converts ReferencedFileID directly to Path(...)
• FileSet.load() checks only (root / file_id).resolve(strict=True) to confirm existence
• FileSet.load() does not verify that the final resolved path is contained within the File-set root
• FileInstance.path returns self.file_set.path / self.node._file_id
• FileSet.copy() uses shutil.copyfile(instance.path, dst)
• FileSet.write() uses Path(instance.path).unlink() and shutil.move(...)

Because there is no containment check such as resolved.relative_to(root.resolve(strict=True)), a malicious DICOMDIR can reference:

• absolute paths such as /etc/passwd
• traversal paths such as ../...
• syntactically conformant file IDs that escape via symlinks

This is not limited to obviously invalid VR input. Even when pydicom emits warnings for invalid ReferencedFileID values, the operation is not blocked. I also confirmed a symlink-based variant using a conformant file ID.

A realistic server-side scenario is:

1. a user uploads a DICOM File-set zip
2. the server loads the uploaded DICOMDIR using FileSet
3. the server re-exports or reorganizes the File-set using FileSet.copy() or FileSet.write()
4. a server-local file referenced by the malicious DICOMDIR is included in the exported result

PoC

Minimal reproduction:

1. Copy a sample File-set that contains a valid DICOMDIR
2. Modify one...