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.
{
"cna_assigner": "VulnCheck",
"cwe_ids": [
"CWE-190",
"CWE-787"
],
"osv_generated_from": "https://github.com/CVEProject/cvelistV5/tree/main/cves/2026/43xxx/CVE-2026-43629.json",
"unresolved_ranges": [
{
"extracted_events": [
{
"introduced": "b4882"
},
{
"last_affected": "b9058"
},
{
"introduced": "0.16.1"
},
{
"last_affected": "0.17.1"
}
],
"source": "AFFECTED_FIELD"
},
{
"extracted_events": [
{
"introduced": "b4882"
},
{
"last_affected": "b9058"
},
{
"introduced": "0.16.1"
},
{
"last_affected": "0.17.1"
}
],
"source": "CPE_FIELD"
},
{
"extracted_events": [
{
"introduced": "b4882"
},
{
"fixed": "b9058"
}
],
"source": "DESCRIPTION"
}
]
}