Ollama's Three-Year RCE Streak: Model Parsing Flaws Persist

Ollama has suffered three remote-code-execution vulnerabilities since 2022, all from parsing untrusted model files. The latest, an out-of-bounds write in MLLAMA metadata parsing, affects all versions before 0.7.0. Attackers can exploit these flaws by uploading malicious models to public registries, enabling code execution on self-hosted runtimes.