Critical Bug Could Expose 300,000 Ollama Deployments to Information Theft

Security researchers are warning that roughly 300,000 deployments of Ollama, a popular open-source tool for running AI models locally, are vulnerable to a critical security flaw dubbed Bleeding Llama. The vulnerability, tracked as CVE-2026-7482 with a severity score of 9.3, allows attackers to steal sensitive information including API keys, prompts, and messages through just three unauthenticated API calls, with no credentials required. The flaw has been patched in Ollama version 0.17.1, and organizations are urged to update immediately and restrict network access to their deployments.