CVE-2026-24221: NVIDIA NVTabular contains a vulnerability where an attacker could cause improper deserialization of untrusted data

Synopsis

Improper deserialization of untrusted data in NVIDIA NVTabular allows a local attacker with a low-privilege account to trigger arbitrary code execution, data tampering, and information disclosure. The vulnerability is reachable locally, requires no victim interaction, and exploits the way NVTabular processes serialized data objects without proper validation. No fix version has been published yet; HarborGuard tracks the advisory and will surface a patched-image rebuild the moment upstream ships a fix.

HarborGuard Coverage

Exploit Conditions

• Network reachabilityNot required

  The attacker needs an existing shell or process on the host; no network exposure is required to reach the vulnerable deserialization path.

• AuthenticationRequired

  Any low-privilege local account is sufficient; no administrative or elevated credentials are needed to trigger the vulnerability.

• Victim interactionNot required

  The exploit does not require any action from another user or a victim; the attacker can trigger deserialization autonomously.

• Attack complexityDetail

  Attack complexity is low, meaning the exploit is reliable and requires no specific race condition, memory layout dependency, or environmental prerequisite beyond local access.

Blast Radius

• A successful attacker executes arbitrary code in the context of the NVTabular process, which can include spawning subprocesses or loading malicious libraries.
• The attacker reads confidential data accessible to the process, such as training datasets, model weights, credentials, or environment variables.
• The attacker modifies or corrupts data processed by NVTabular, including persisted feature-engineering outputs, cached datasets, or pipeline artifacts.
• The exploit can render the NVTabular service or worker process unavailable by crashing or hijacking it, disrupting any dependent training or inference pipeline.