CVE-2026-76843: Deserialization of Untrusted Data in flairNLP flair

The vulnerability resides in flairNLP flair versions 0.15.0 and 0.15.1 in the flair/models/clustering.py file. The ClusteringModel.load static method uses pickle.loads(joblib.load(str(model_file))) to load model files, which can execute arbitrary Python code during deserialization. Although clustering support was dropped from the documented API in 0.15.0, the clustering.py module remains in the distributed package and can be imported directly, making the vulnerability still exploitable. This is the same code path as CVE-2024-10073, but the previous fix was ineffective because the vulnerable code was not removed from the distributed artifact. No official remediation or patch is currently documented.

An attacker who can supply a malicious model file to be loaded by the ClusteringModel.load method can execute arbitrary Python code with the privileges of the process loading the model. This can lead to full compromise of the environment running the flair library. The vulnerability is local vector (AV:L) with low attack complexity but requires user interaction. The impact includes high confidentiality, integrity, and availability consequences due to arbitrary code execution.

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. Until an official fix is released, avoid loading untrusted or attacker-supplied model files using the ClusteringModel.load method. Consider removing or restricting access to the flair.models.clustering module if clustering functionality is not required. Monitor official flairNLP channels for updates and patches addressing this issue.