CVE-2026-43629

High [CVE-2026-43629] Arbitrary code execution via crafted KV cache state files

CVE-2026-43629Source published Aug 6, 2026Source updated Aug 6, 2026

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. An attacker with local write access to specific directories can exploit a heap buffer overflow vulnerability. By crafting malicious data files, the attacker can cause the system to write…

Affected products in this advisory

• Red Hat Enterprise Linux AI (RHEL AI) 3

Source-reported affected versions

Affected-version details not available in this record.

Source-reported fixed versions

No fixed-version detail extracted. This does not mean no fix exists.

Mitigation guidance

• Restrict write access to the `slot_save_path` directory used by llama.cpp. Ensure that only trusted processes and users have permissions to write to this location. This prevents an attacker from introducing malicious KV cache state files that could trigger the heap buffer overflow. For deployments, consider running llama.cpp with the least necessary privileges and within a sandboxed environment to further limit the impact of potential file system compromises.