ENHANCING CLOUD SECURITY IN AI/ML: THE LITTLE PICKLE STORY

Summary from the TLDR InfoSec newsletter

Python pickles are a serialization format that is commonly used in AI/ML workflows in Python. Pickles that come from unsafe sources can result in arbitrary code execution due to invalidated object loading, corrupt models, exposure of sensitive data, or lead to sensitive information disclosure due to a lack of encryption. One strategy for scanning pickle files using static code analysis is to generate a list of all Python libraries that are used in your environment, validate any pre-built rules in existing static code analysis tooling, and write custom rules in tools like Semgrep to identify pickling and unpickling methods.