Meta fairseq Library Suffers Critical RCE Vulnerability via Deserialization Flaw
A remote code execution vulnerability in Meta's fairseq library, disclosed by Zero Day Initiative, allows attackers to compromise systems by exploiting a deserialization flaw.
Zero Day Initiative (ZDI) has disclosed a critical remote code execution (RCE) vulnerability, tracked as ZDI-26-944, affecting Meta's fairseq library. The flaw resides within the load_ema_from_checkpoint method, where improper validation of user-supplied data leads to the deserialization of untrusted data. This vulnerability carries a CVSS score of 7.8, indicating a high severity.
Exploitation of this vulnerability requires user interaction. Attackers can leverage this by tricking a victim into visiting a malicious webpage or opening a specially crafted file. Once triggered, the vulnerability allows remote attackers to execute arbitrary code within the context of the current process on affected installations of Meta fairseq.
The disclosure timeline reveals a diligent reporting process. ZDI initially reported the vulnerability to Meta on March 4, 2026. Meta acknowledged the report the following day and communicated that the issue was under investigation by March 17. However, Meta later informed ZDI that the fairseq repository was no longer maintained and had been archived on March 20, 2024. Notably, public records indicate the repository was actually archived on March 20, 2026, after ZDI's report was submitted.
Given that the fairseq repository has been archived and is no longer maintained by Meta, a direct patch is unlikely to be provided. The only effective mitigation strategy recommended by ZDI is to restrict interaction with the product. This means users should avoid opening files or visiting links associated with the fairseq library from untrusted sources.
The fairseq library is a sequence-to-sequence modeling toolkit developed by Meta AI, widely used in natural language processing research and development. Its use in machine learning workflows means that vulnerabilities within it can have significant implications for organizations relying on these tools for training and deploying AI models.
The archiving of the fairseq repository raises concerns about the long-term security posture of projects that depend on it. While Meta has moved on to other projects, the vulnerability highlights the ongoing risk posed by unmaintained software libraries, which can become targets for attackers seeking to exploit known but unpatched flaws.
This vulnerability, discovered by Bongeun Koo, underscores the persistent threat of deserialization flaws in software libraries, particularly those handling complex data structures like model checkpoints. The reliance on user interaction for exploitation, while a mitigating factor, does not diminish the critical nature of the RCE capability it enables.
As the fairseq library is no longer actively maintained, users are strongly advised to assess their exposure and consider migrating to actively supported alternatives or implementing strict access controls to prevent potential exploitation of this vulnerability.
This advisory, ZDI-26-943, details a deserialization vulnerability in Meta's fairseq library's load_checkpoint_to_cpu method. The vulnerability requires user interaction, such as opening a crafted file or visiting a malicious webpage, and allows for remote code execution in the context of the current process. While the fairseq repository was archived on March 20, 2026, indicating it is no longer actively maintained, the Zero Day Initiative has published this advisory due to the potential for exploitation.
This advisory, ZDI-26-939, details a zero-day vulnerability in Meta's fairseq2 library, specifically within the FileCheckpointManager's _load_state_dict method. The flaw allows for deserialization of untrusted data, leading to remote code execution. Notably, the fairseq repository has been archived and is no longer maintained by Meta, meaning a patch is unlikely, and users are advised to restrict interaction with the product.
The Zero Day Initiative (ZDI) has published advisory ZDI-26-937 detailing a deserialization vulnerability within Meta's fairseq library. This advisory highlights that the fairseq repository is no longer maintained and has been archived by Meta, meaning no official patches will be released. The vulnerability, which allows for remote code execution with user interaction, was reported to the vendor on March 26, 2026, and is now publicly disclosed as a zero-day.
This advisory from Zero Day Initiative (ZDI-26-921) details a remote code execution vulnerability in Meta's fairseq TransformerLanguageModel, specifically within its YAML file parsing. The vulnerability requires user interaction, where an attacker tricks a user into opening a malicious file. Notably, Meta has indicated that the fairseq repository is no longer maintained and has been archived, meaning a patch is unlikely.
This advisory from Zero Day Initiative details a deserialization vulnerability within Meta's fairseq library's dump_km_label module. The vulnerability, assigned a CVSS score of 7.8, allows for remote code execution when a user interacts with a malicious page or file. Notably, Meta has indicated that the fairseq repository is no longer maintained and has been archived, leaving affected installations without a vendor patch.