A critical authentication bypass in the NVIDIA Triton Inference Server allows unauthenticated attackers to gain full control, leading to potential code execution and data tampering.

The vulnerability is an authentication bypass using an alternate path or channel (CWE-288). It allows an unauthenticated attacker to bypass security controls, potentially resulting in unauthorized code execution, privilege escalation, or full system disruption.

The ability to bypass authentication on an inference server poses a significant threat to AI/ML infrastructure. Exploitation could allow an attacker to exfiltrate proprietary models, corrupt training data, or utilize compute resources for unauthorized tasks. With a CVSS score of 9.8, this represents a critical threat to the integrity and availability of critical backend systems.

Immediate Action: Upgrade to NVIDIA Triton Inference Server version r26.03 or later.

Proactive Monitoring: Monitor server logs for unusual API traffic, unauthorized access attempts, or spikes in resource utilization that may indicate exploitation.

Compensating Controls: Implement strict network-level access controls, such as VPNs or mTLS, to ensure that the inference server is not exposed to untrusted networks.

Public Exploit Available: Yes — a public proof-of-concept repository exists on GitHub.

Given the critical nature of this flaw and the presence of public exploit code, organizations running NVIDIA Triton Inference Server must prioritize patching. Failure to update to version r26.03 leaves the infrastructure exposed to full administrative takeover by remote attackers.