Ollama, Nvidia Flaws Put AI Infrastructure at Risk

Security researchers identified multiple vulnerabilities in AI infrastructure products, including Ollama and NVIDIA’s Triton Inference Server, with one flaw capable of remote code execution.

According to security firm Fuzzinglabs, four vulnerabilities were found in Ollama and one in Triton, including CVE-2024-12886 (a denial-of-service flaw) and CVE-2025-51471 (an authentication bypass), plus CVE-2025-48889 (an arbitrary file copy) and a heap overflow bug without a CVE yet; CVE-2024-12886 carries a CVSS score of 7.5.

In Triton Server, researchers uncovered a command-injection vulnerability in the model configuration pipeline, which Ventuzelo says was easy to exploit and could have led to remote code execution without prior authentication. The flaws were disclosed ahead of Black Hat Europe 2025, where the findings are set to be presented on 10 December 2025 as part of the event in London.

Though Ollama and NVIDIA have since fixed the issues, the report underscores how AI security research is increasingly focusing on the infrastructure that runs models, not just the models themselves.