Executive brief
MosaicML Composer is an open-source library used for training machine learning models. A security vulnerability in how the software handles 'checkpoint' files could allow an attacker to execute malicious code on a user's system. To exploit this, an attacker would need to trick a user into opening a specially crafted malicious file or visiting a compromised webpage. Successful exploitation could lead to a full system compromise or unauthorized access to sensitive training data.
Technical details
A deserialization vulnerability (CWE-502) exists in MosaicML Composer within the checkpoint parsing logic. The root cause is the use of the Python 'pickle' module to process user-supplied checkpoint data without adequate validation or restricted global access. An attacker can craft a malicious serialized object that, when unpickled, executes arbitrary commands. While categorized as remote code execution, the attack vector is local (AV:L) requiring user interaction (UI:R), such as loading a malicious checkpoint file. The vulnerability has been addressed in a commit that implements a restricted 'Unpickler' using a whitelist of safe globals (_RESUMPTION_SAFE_GLOBALS).
Affected products
- MosaicML Composer 0.32.1
Timeline
- 2026-02-12: disclosed: Vulnerability reported to vendor
- 2026-04-29: patched: Fix committed to GitHub repository
- 2026-06-24: advisory: Coordinated public release of advisory