Database/AI/ML frameworks & serving
NVIDIA Megatron-LM: unsafe deserialization of untrusted data allows code execution
Impact
Megatron-LM deserializes untrusted data in a code path the record does not name. A successful exploit runs code as the training process, which holds the job's GPUs, data and credentials. NVIDIA rates the impact as code execution, information disclosure and data tampering. The same bulletin also covers separate dataset-loading and checkpoint-loading flaws (CVE-2026-100968, CVE-2026-65132).
Who can reach it
Local, low privilege: a user who can supply input that a Megatron-LM process deserializes. The record does not say which input. No user interaction needed.
What to do
Upgrade Megatron-LM to the fixed release named in NVIDIA bulletin 5926 and restart the training jobs that use it. The record does not give a fixed version.
References
Related entries
- NVIDIA Megatron-LM: A further script-level code-injection path, filed under the same class as the 2025 setCVE-2026-24149 · NVIDIA Megatron-LMHigh
- NVIDIA Megatron-LM: Checkpoint loading reaches remote code execution when a user loads a crafted checkpointCVE-2026-24150 · NVIDIA Megatron-LMHigh
- NVIDIA Megatron-LM: The inferencing path reaches remote code execution on crafted inputCVE-2026-24151 · NVIDIA Megatron-LMHigh
- NVIDIA Megatron-LM: A second checkpoint-loading remote code execution pathCVE-2026-24152 · NVIDIA Megatron-LMHigh
- NVIDIA Megatron-LM: A code-injection flaw in the tools component executes attacker-controlled code inside the trainingCVE-2025-23305 · NVIDIA Megatron-LMHigh
- NVIDIA Megatron-LM: megatron/training/arguments.py injects code from malicious inputCVE-2025-23306 · NVIDIA Megatron-LMHigh
This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.