NVIDIA has disclosed critical security vulnerabilities in its Megatron LM software that enable attackers to inject malicious code through compromised files, prompting an immediate security update release.
The company issued Security Bulletin on June 24, 2025, identifying two high-severity vulnerabilities (CVE-2025-23264 and CVE-2025-23265) that affect all versions of Megatron LM prior to version 0.12.0, with both carrying CVSS scores of 7.8.
Users are strongly advised to upgrade to version 0.12.1 or later to protect their systems from potential code execution, privilege escalation, information disclosure, and data tampering attacks.
NVIDIA’s security team has identified two distinct but related vulnerabilities within the Python components of Megatron LM, both classified under CWE-94 (code injection).
The vulnerabilities, designated as CVE-2025-23264 and CVE-2025-23265, share identical attack vectors and impact profiles, suggesting they may stem from similar underlying code weaknesses.
Both vulnerabilities allow attackers to exploit the system by providing malicious files that can be processed by the vulnerable Python components.
The attack vector classification reveals that these vulnerabilities require local access (AV:L) with low attack complexity (AC:L) and low privileges required (PR:L), while requiring no user interaction (UI:N).
This combination makes the vulnerabilities particularly concerning for environments where multiple users have system access or where automated processes handle file inputs.
The scope remains unchanged (S:U), but the impact on confidentiality, integrity, and availability is rated as high across all three categories.
NVIDIA Megatron LM Vulnerabilities
The security bulletin indicates that these vulnerabilities affect Megatron LM installations across all platforms, presenting several critical concerns:
- Universal Platform Impact: The vulnerabilities affect all operating systems that support Megatron LM, including Linux, Windows, and other supported platforms, making the issue universally critical for organizations utilizing this large language model framework.
- No Deployment Exemptions: The comprehensive nature of the affected systems means that no deployment environment is exempt from potential exploitation, regardless of the underlying infrastructure or configuration setup.
- Variable Risk Assessment: NVIDIA’s risk assessment acknowledges that the severity may vary depending on specific deployment configurations, stating that their evaluation represents “an average of risk across a diverse set of installed systems.”
- Customized Risk Evaluation Recommended: Organizations are encouraged to conduct their own risk assessments based on their particular implementation details, user access patterns, and security controls already in place to determine their specific exposure level.
Immediate Update
The security update addresses both identified vulnerabilities through the release of Megatron LM version 0.12.1, available through the official NVIDIA GitHub repository.
Users must clone or update their installations to this version or later to ensure protection against potential attacks.
The update process involves accessing the Releases section of the NVIDIA/Megatron-LM repository on GitHub.
NVIDIA has credited security researchers Yu Rong and Hao Fan for responsibly disclosing these vulnerabilities, following standard coordinated disclosure practices.
The company emphasizes that earlier software branch releases supporting the product are also affected, requiring users on legacy branches to upgrade to the latest available branch release.
Organizations using Megatron LM in production environments should prioritize this update, particularly those handling sensitive data or operating in multi-user environments where the attack vector could be more readily exploited.
NVIDIA continues to monitor for additional security issues and maintains its Product Security page for ongoing updates and vulnerability reporting.
Find this Story Interesting! Follow us on LinkedIn and X to Get More Instant Updates.
.webp?w=356&resize=356,220&ssl=1)




