Impact
A path traversal weakness exists in the load_state_dict_into_model function of the lxtGH OMG‑Seg MoE Shard Checkpoint Loader, triggered when an attacker manipulates the weight_map argument. This flaw allows the construction of file paths that escape the intended directory boundary, which could lead to arbitrary file reads or writes. Because the loader processes externally supplied weight_map data, an attacker can use the vulnerability to retrieve sensitive configuration files or overwrite critical files, enabling possible code execution or privilege escalation if executable files are overwritten or accessed.
Affected Systems
The affected product is lxtGH's OMG‑Seg. The vulnerability was present in all releases up to and including commit f3b26b609b. Since the project does not use semantic versioning, specific version numbers are not available; therefore any build that contains that commit or earlier may be impacted. The component is the MoE Shard Checkpoint Loader located in omg_llava/xtuner/utils/handle_moe_load_and_save.py. The product is hosted on GitHub and the issue has been reported but the maintainer has not responded. No official patch or workaround is currently available.
Risk and Exploitability
The CVSS v3 base score of 5.3 indicates a medium impact for the affected system. EPSS data is not reported, and the vulnerability is not listed in the CISA KEV catalog, yet a public exploit is available, demonstrating that remote exploitation is feasible. The attack vector is inferred to be network-based because malicious weight_map payloads can be supplied remotely to the loader process. An adversary controlling the weight_map can execute path traversal to read or modify files on the host system, potentially impacting confidentiality, integrity, and availability.
OpenCVE Enrichment