Overview
- Wiz researchers disclosed three high-severity CVEs in Triton’s Python backend on August 4 that can be chained by unauthenticated actors for remote code execution.
- CVE-2025-23319 (CVSS 8.1), CVE-2025-23320 (CVSS 7.5) and CVE-2025-23334 (CVSS 5.9) enable out-of-bounds write, shared memory abuse and read vulnerabilities.
- A successful exploit chain could result in full server compromise with data tampering, denial of service and theft of proprietary AI models.
- NVIDIA’s August bulletin and Triton version 25.07 also address three additional critical bugs (CVE-2025-23310, CVE-2025-23311 and CVE-2025-23317) to strengthen AI deployment security.
- There are no reports of active exploitation in the wild and organizations are urged to apply the latest updates immediately to protect their AI infrastructure.