1  info:
 2    name: llama-cpp
 3    cve: CVE-2026-33298
 4    summary: "llama.cpp: integer overflow in ggml_nbytes allows heap buffer overflow via crafted GGUF file"
 5    details: >-
 6      llama.cpp prior to build b7824 contains an integer overflow vulnerability in the
 7      ggml_nbytes function. An attacker can craft a malicious GGUF file with specific tensor
 8      dimensions to cause ggml_nbytes to return a significantly smaller size than required
 9      (e.g., 4MB instead of Exabytes), bypassing memory validation. This leads to a
10      heap-based buffer overflow when the application subsequently processes the tensor,
11      potentially resulting in Remote Code Execution (RCE) via memory corruption.
12      The vulnerability requires user interaction (opening a crafted GGUF file) and has
13      a local attack vector with CVSS score 7.8 (HIGH). No public PoC has been identified
14      at the time of reporting. CWE classifications: CWE-190 (Integer Overflow), CWE-122
15      (Heap-based Buffer Overflow).
16    cvss: "CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H"
17    severity: HIGH
18    security_advise: >-
19      Upgrade llama.cpp to build b7824 or later. Do not load GGUF files from untrusted sources.
20      The fix is available at: https://github.com/ggml-org/llama.cpp/releases/tag/b7824
21    references:
22      - https://github.com/ggml-org/llama.cpp/security/advisories/GHSA-96jg-mvhq-q7q7
23      - https://github.com/ggml-org/llama.cpp/releases/tag/b7824
24      - https://nvd.nist.gov/vuln/detail/CVE-2026-33298
25  rule: ""
26  references:
27    - https://github.com/ggml-org/llama.cpp/security/advisories/GHSA-96jg-mvhq-q7q7
28    - https://nvd.nist.gov/vuln/detail/CVE-2026-33298