Lethal Language Models - From Bit Flip to RCE in Ollama

Explore critical security vulnerabilities in the Ollama large language model platform through this 28-minute conference talk that demonstrates how seemingly minor bit flip errors can escalate into remote code execution attacks. Learn about the technical mechanisms behind these exploits, understand the attack vectors that transform innocent-looking model corruptions into serious security breaches, and discover the implications for AI system security. Examine real-world examples of how attackers can manipulate language models to achieve unauthorized system access, analyze the underlying causes of these vulnerabilities in Ollama's architecture, and gain insights into defensive strategies for protecting AI deployments from similar attacks.