NVIDIA Discloses Critical Triton Authentication Bypass Vulnerability

NVIDIA disclosed a critical authentication-bypass vulnerability in its Triton Inference Server, tracked as CVE-2026-24207 with a CVSS 3.1 score of 9.8, exploitable remotely without credentials or user interaction, according to GBHackers and NVIDIA's own May 2026 security bulletin, verified via NVD. Successful exploitation could allow code execution, privilege escalation, data tampering, denial-of-service, or exposure of sensitive AI-workload data, since the flaw (CWE-288) affects all Triton and DALI Backend versions before r26.03, per NVD's confirmed affected-version data. The same bulletin also discloses a related authentication-bypass flaw, CVE-2026-24206, scored lower by NVIDIA (7.3) though NVD's own independent assessment also rates it 9.8; other outlets report additional related CVEs in the same release. Organizations running Triton pipelines should update to r26.03 or later immediately.
The most important detail practitioners may miss is that this is not a single-CVE patch: NVIDIA's May 2026 Triton bulletin bundles CVE-2026-24207 (the critical 9.8 flaw enabling code execution) with a related authentication-bypass CVE-2026-24206 and, per multiple security outlets, several other CVEs affecting Triton and its DALI Backend, all fixed in the same r26.03 release.
What happened
NVIDIA disclosed a critical authentication-bypass vulnerability in its Triton Inference Server, tracked as CVE-2026-24207, with a CVSS 3.1 base score of 9.8 (critical), according to GBHackers and NVIDIA's May 2026 security bulletin. This was independently confirmed via NVD, which lists NVIDIA as the CNA and records the same 9.8 score and vector (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H). The flaw is classified under CWE-288 (Authentication Bypass Using an Alternate Path or Channel) and is exploitable remotely, without authentication or user interaction; successful exploitation could lead to code execution, privilege escalation, data tampering, denial of service, or information disclosure. NVD's affected-configuration data confirms all Triton Inference Server versions before r26.03 are impacted, and NVIDIA has released r26.03 as the fix. The same bulletin also discloses CVE-2026-24206, a related authentication-bypass flaw (also CWE-288, same affected versions) that NVIDIA's own CNA scored lower, at 7.3 HIGH, because it lacks the code-execution impact; notably, NVD's own independent assessment scores CVE-2026-24206 at 9.8 as well, a discrepancy between NIST's and NVIDIA's severity ratings. Other security outlets, including securityonline.info and cyberpress.org, report the same bulletin covers several additional related CVEs affecting Triton and DALI Backend components, all addressed in r26.03.
Technical context
The vulnerability stems from improper authentication controls in Triton, according to GBHackers' review of NVIDIA's bulletin, allowing attackers to circumvent authentication mechanisms via the CWE-288 pathway. Inference servers like Triton sit at a high-value point in production AI stacks: they receive networked requests, execute model code paths, and often host or access proprietary model artifacts and sensitive data, so an unauthenticated remote-code-execution path on this infrastructure carries outsized risk for GPU-accelerated inference deployments in cloud and edge environments.
For practitioners
Update Triton Inference Server (and the DALI Backend, if deployed) to r26.03 or later immediately; this single release addresses CVE-2026-24207 alongside the related CVE-2026-24206 and other CVEs disclosed in the same bulletin, so a partial or selective patch is not sufficient. In the interim, or for systems that cannot be patched immediately, restrict network access to inference endpoints with allowlists, place servers behind authentication proxies, and monitor for unusual requests to inference endpoints or unexpected process launches tied to Triton.
What to watch
- •The official NVIDIA security bulletin and Triton GitHub repository for any updates to the affected-CVE list or fix guidance.
- •Public proof-of-concept exploit disclosures on vulnerability feeds, given the low exploitation complexity.
- •Whether NVD's own CVSS enrichment (currently marked "not yet provided" for CVE-2026-24207 itself) converges with or diverges from NVIDIA's CNA scoring, as it already has for the companion CVE-2026-24206.
Key Points
- 1NVIDIA's Triton Inference Server has a critical, remotely exploitable authentication-bypass flaw (CVE-2026-24207, CVSS 9.8) enabling code execution.
- 2The same May 2026 bulletin patches a related authentication-bypass CVE-2026-24206 and other CVEs, all fixed together in version r26.03.
- 3NVD's own severity assessment for the companion CVE-2026-24206 diverges sharply from NVIDIA's, rating it 9.8 versus NVIDIA's 7.3.
Scoring Rationale
Critical (CVSS 9.8), remotely exploitable, unauthenticated code-execution flaw in widely-deployed production AI inference infrastructure, independently confirmed via NVD's authoritative CVE record. Score reflects real operational risk to GPU-accelerated inference deployments; kept just below 'industry-shaking' since this is a standard (if severe) vendor disclosure-and-patch cycle with a fix already available, not an actively-exploited zero-day.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
