CVE-2026-43629

Summary

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.

Affected Software

VendorProductVersion RangeStatus
ggml-orgllama.cppb4882 <= b9058affected
ggml-orgllama.cpp0.16.1 <= 0.17.1affected

Weaknesses

  • CWE-787: CWE-787 Out-of-bounds Write
  • CWE-190: CWE-190 Integer Overflow or Wraparound

References