VYPR
advisoryPublished Nov 13, 2026· Updated Nov 21, 2026· 3 sources

Zero-Day Vulnerability in llama.cpp Allows Remote Code Execution via Malicious GGUF Files

A critical remote code execution vulnerability (CVE-2026-126976) has been discovered in llama.cpp's multimodal GGUF file parsing, allowing attackers to compromise systems by tricking users into opening malicious files.

A significant security flaw has been identified in the popular open-source large language model inference library, llama.cpp. The vulnerability, tracked as CVE-2026-126976 and disclosed by the Zero Day Initiative (ZDI), allows remote attackers to execute arbitrary code on affected systems.

Exploitation of this vulnerability requires user interaction, specifically that the target user must open a malicious file. The flaw resides within the parsing logic for audio multimodal projector GGUF files. Attackers can leverage a lack of proper validation of user-supplied data, which leads to an integer overflow condition before a buffer is allocated.

This integer overflow can be exploited by an attacker to overwrite memory and ultimately achieve arbitrary code execution within the context of the current process. The Zero Day Initiative has assigned this vulnerability a CVSS score of 7.0, highlighting its considerable security risk. The vulnerability affects installations of llama.cpp that process these specific GGUF file types.

The developers of llama.cpp have already released a patch to address this critical vulnerability. The fix is available in their latest commit on GitHub, which details the necessary code changes to prevent the integer overflow and subsequent code execution. Users are strongly advised to update their llama.cpp installations to the patched version as soon as possible to mitigate the risk.

The disclosure timeline indicates that the vulnerability was initially reported to the vendor on June 18, 2026. Following a coordinated public release of the advisory on November 13, 2026, the advisory was updated on the same day. This timeline suggests a diligent process between the vulnerability reporter and the developers to ensure a secure resolution.

The vulnerability was discovered and reported by a security researcher known as 'zent'. The Zero Day Initiative's advisory provides detailed technical information for security professionals and system administrators to understand the threat and implement necessary countermeasures. The prompt and detailed nature of the advisory underscores the importance of timely patching and security awareness.

This discovery highlights the ongoing security challenges associated with complex software libraries that handle diverse file formats. As AI models and their associated tooling become more prevalent, the attack surface expands, making robust input validation and secure parsing mechanisms paramount. The llama.cpp team's swift response to this vulnerability is a positive indicator of their commitment to security.

Organizations and individuals utilizing llama.cpp should prioritize applying the available security update. Failure to do so could leave them susceptible to remote code execution attacks, potentially leading to system compromise, data theft, or further malicious activity. Staying informed about such vulnerabilities and promptly applying patches is a critical aspect of maintaining a secure computing environment.

This advisory updates the previous report by providing the specific CVE identifier, CVE-2026-126975, and a CVSS score of 7.8 for the vulnerability. It also details that the flaw is a heap-based buffer overflow within the GGUF file parsing component, which can be triggered by user interaction through opening a malicious file or visiting a compromised webpage.

This advisory details a specific integer overflow vulnerability within the GGUF file parsing component of llama.cpp, assigned CVE-2026-126974. The vulnerability requires user interaction, such as opening a crafted file or visiting a malicious webpage, to achieve arbitrary code execution in the context of the current process. Zero Day Initiative has assigned this vulnerability a CVSS score of 7.8.

Synthesized by Vypr AI