Are Your Machine Learning Frameworks Safe from Exploitation?

The reliance on machine learning (ML) frameworks by organizations for various applications has grown exponentially, raising numerous questions about their security. Recent disclosures by JFrog’s researchers have spotlighted significant vulnerabilities in popular open-source ML frameworks like MLflow, PyTorch, and MLeap. Unlike previous concerns, which mainly revolved around server-side issues, these new flaws make it possible for attackers to exploit ML clients through libraries designed to manage secure model formats such as Safetensors. The potential impact of these vulnerabilities is staggering, as exploiting an ML client can enable attackers to move laterally within an organization and access sensitive information, including model registry credentials. For organizations leveraging these ML frameworks, comprehending the nature and potential risks of these vulnerabilities is essential to prevent catastrophic security breaches.

Key Vulnerabilities in Popular ML Frameworks

Central to the security concerns are several critical vulnerabilities identified across different ML frameworks. Among these is CVE-2024-27132, an issue in MLflow where insufficient sanitization opens the door to cross-site scripting (XSS) attacks, potentially leading to client-side remote code execution (RCE). Adding to these concerns is CVE-2024-6960 in ##O, which reveals an unsafe deserialization problem capable of resulting in RCE when an untrusted ML model is imported. These flaws highlight the significant risks associated with trust boundaries in ML frameworks, where injecting malicious models can lead to extensive system compromise and unauthorized data access.

Additionally, in PyTorch, the TorchScript feature is compromised by a path traversal issue that could cause denial-of-service (DoS) or the overwriting of arbitrary files. Such vulnerabilities can potentially compromise critical system files, leading to severe disruptions or unauthorized access. MLeap is not safe from these issues either; CVE-2023-5245 identifies a path traversal issue causing a Zip Slip vulnerability when loading a saved model in a zipped format. This flaw allows for arbitrary file overwriting and possible code execution, opening avenues for malicious attacks that could cripple essential ML operations.

Caution Is Necessary Even with Trusted Sources

Given these vulnerabilities, the importance of cautious handling of machine learning models cannot be overstated. Even models from reliable sources like Safetensors can pose significant risks. Organizations must verify the integrity of the ML models they use, ensuring they don’t unintentionally introduce potential backdoors. Shachar Menashe, JFrog’s VP of Security Research, highlights the dual nature of AI and ML tools: while they offer significant innovation potential, they can become harmful attack vectors if untrusted models are loaded. He advocates for a systematic, careful approach to using these models, stressing the need for security protocols that guard against remote code execution and other malicious exploits.

To mitigate these risks, organizations should implement stringent verification processes for all ML models, regardless of their origin. Investing in robust security measures, such as regular audits and checks, helps identify and mitigate potential threats before they cause damage. Additionally, maintaining a knowledgeable IT team updated with the latest security practices can significantly reduce the likelihood of successful attacks. Lessons from these vulnerabilities remind us of the constantly evolving security threats in ML technologies. To sustain ML benefits while minimizing risks, consistent vigilance and proactive security measures are essential.

Explore more