Status: UPDATED
| Advisory ID: CVE-2026-43629
Key Details
| CVE | CVE-2026-43629 |
| CVSS Score / Version | 8.1 (High) / CVSS v3.1 |
| Updated | 2026-09-21 |
| CVSS Vector | CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H |
| CVSS Prose | attack vector is network; attack complexity is high; privileges required is none; user interaction is none; scope is unchanged; confidentiality impact is high; integrity impact is high; availability impact is high. |
| Affected products | ggml llama.cpp |
| Classified as | CWE-190 (Integer Overflow or Wraparound) |
Affected Products, Subsystems & Sectors
| Subsystems | General OT |
| Sectors | Multiple |
What to Know
llama.cpp builds b4882 through b9058 contain a heap buffer overflow vulnerability in the KV cache state restore path where the state_read_data() function computes write size without overflow checking, allowing attackers with write access to the slot_save_path directory to corrupt heap memory. Attackers can craft malicious state files where cell_count multiplication overflows or exceeds tensor buffer allocation to write attacker-controlled bytes past buffer boundaries, potentially resulting in heap metadata corruption, model weight corruption, or arbitrary code execution via function pointer overwrite. (NVD)
What to Do
Monitor ggml's web page for any future patch releases.
References