The mamba language model framework through version 2.2.6 is vulnerable to insecure deserialization (CWE-502) due to the use of torch.load() without the weights_only=True parameter in the MambaLMHeadModel.from_pretrained() method. This allows an attacker to craft a malicious model repository on HuggingFace Hub that, when loaded by a victim, results in arbitrary code execution via pickle deserialization. The vulnerability has a CVSS 3.1 base score of 9.8, indicating critical severity with network attack vector, no required privileges or user interaction, and full confidentiality, integrity, and availability impact. No official patch or remediation guidance is currently available, and no known exploits are reported in the wild.

Successful exploitation allows an attacker to execute arbitrary code on the victim system running the mamba framework by tricking the victim into loading a malicious pre-trained model from HuggingFace Hub. This compromises confidentiality, integrity, and availability of the affected system. The vulnerability is remotely exploitable without authentication or user interaction.

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. Until an official fix is available, users should avoid loading pre-trained models from untrusted or unknown HuggingFace repositories. Consider manual inspection or sandboxing when loading external models to reduce risk.