AI Stack Alert High CVE-2026-65110 For AI Eng, AppSec, SOC / IR AI-assisted
NVIDIA Megatron-LM checkpoint deserialization flaw allows remote code execution (CVE-2026-65110)
A deserialization of untrusted data flaw exists in how NVIDIA Megatron-LM parses checkpoints. An attacker can execute arbitrary code in the context of the current process if the target opens a malicious file. The issue is fixed in Megatron-LM version 0.17.1.
Why it matters
Teams that load Megatron-LM checkpoints from untrusted or shared sources may run attacker code with the privileges of the training or inference process. Exposed model training environments and their credentials could be affected.
Do this
- Upgrade Megatron-LM to version 0.17.1 or later.
- Load checkpoints only from trusted, verified sources until patched.
- Review NVIDIA product security bulletin 5881 for affected deployments and scope.
- Inventory hosts running Megatron-LM and identify where checkpoint files are ingested.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.23%
Signals
- NoCISA KEV
- 0.23%EPSS
- 36/100Priority
- 80%Confidence
Evidence
“The specific flaw exists within the parsing of checkpoints. The issue results from the lack of proper validation of user-supplied data, which can result in deserialization of untrusted data.”
“Fixed in Megatron-LM version 0.17.1”
Sources
- Zero Day Initiative: published advisories
- Zero Day Initiative: published advisories
- Zero Day Initiative: published advisories
- Zero Day Initiative: published advisories
- Zero Day Initiative: published advisories
- Zero Day Initiative: published advisories
- Zero Day Initiative: published advisories
- Zero Day Initiative: published advisories
- Zero Day Initiative: published advisories
- Zero Day Initiative: published advisories
- Zero Day Initiative: published advisories
- Zero Day Initiative: published advisories
Affected / related
NVIDIA Megatron-LM, Megatron-LM
Published 2026-11-10 06:00 UTC · updated 2026-11-12 17:35 UTC · AI-assisted summary, checked against quoted sources · How we score