Unpickling PyTorch: Keeping Malicious AI Out of the Enterprise

No ratings

Presented at RSAC 2025 by

PyTorch is a go-to framework for organizations building their own LLMs. But, how can anyone be sure they aren’t accidentally running malicious code? Existing tools that examine the unpickling process produce false positives, posing more questions than answers. This session will demonstrate a new method for developers to identify and extract malicious code before it runs, keeping AI models safe.