Three vulnerabilities in Ollama, a widely deployed open-source AI framework, expose organizations running local large language models to memory theft and persistent code execution. The most severe flaw allows unauthenticated remote attackers to read heap memory from Ollama servers, potentially leaking API keys, credentials, and conversation data from AI workloads. Two Windows-specific vulnerabilities remain unpatched and can be chained to establish persistent code execution, threatening any organization that has deployed Ollama on Windows infrastructure.