Intel has recently disclosed a critical vulnerability in its Intel Neural Compressor software used for AI model compression. The bug, identified as CVE-2024-22476, allows unauthenticated attackers to execute arbitrary code on Intel systems running affected versions of the software. This vulnerability is considered the most severe among the 41 security advisories issued by Intel this week.
The vulnerability, traced back to improper input validation, has been assigned a maximum CVSS score of 10 due to its remote exploitability with low complexity and its high impact on data confidentiality, integrity, and availability. Attackers do not need special privileges or user interaction to exploit this flaw. The affected versions of Intel Neural Compressor are those prior to 2.5.0, and Intel strongly recommends organizations to upgrade to version 2.5.0 or later to mitigate the risk.
Intel Neural Compressor is an open-source Python library designed to compress and optimize deep learning models for various applications such as computer vision, natural language processing, and recommendation systems. The compression techniques involve neural network pruning, memory reduction through process call quantization, and distillation of larger models into smaller ones with comparable performance. The aim of AI model compression technology is to facilitate the deployment of AI applications on different hardware devices, including those with limited computational power like mobile devices.
CVE-2024-22476 is just one of two vulnerabilities in Intel’s Neural Compressor software disclosed and addressed by the company this week. The other vulnerability, CVE-2024-21792, is a time-of-check-time-of-use (TOCTOU) flaw that could lead to information disclosure, but it presents a moderate risk as it requires an attacker to already have local, authenticated access to a vulnerable system.
In addition to the vulnerabilities in the Neural Compressor software, Intel also revealed five high-severity privilege escalation vulnerabilities in its UEFI firmware for server products. These vulnerabilities, categorized as input validation flaws with severity scores ranging from 7.2 to 7.5 on the CVSS scale, highlight the importance of addressing security risks across various components and infrastructure supporting AI products and platforms.
The emergence of vulnerabilities in AI software components underscores the expanding attack surface that AI tools are creating within enterprise organizations. While much attention has been focused on the risks associated with large language models and LLM-enabled chatbots, recent research has unveiled vulnerabilities such as model manipulation and jailbreaking that threaten the security of AI applications.
Researchers have also identified potential cyber risks in core software components and infrastructure used in AI development and deployment processes. Weaknesses in widely used platforms like HuggingFace have been exploited by attackers to tamper with models or upload malicious ones. A recent study commissioned by the UK government highlighted the need for comprehensive threat modeling, secure authentication, and input validation to safeguard AI technologies at every stage of their lifecycle.
Overall, the disclosure of vulnerabilities like CVE-2024-22476 underscores the critical importance of addressing security concerns in AI software and infrastructure to prevent exploitation and protect sensitive data. Businesses and organizations leveraging AI technologies must prioritize security measures to mitigate the risks posed by evolving threats in the AI landscape.