Aim Labs proposes a sandbox-execution-based dynamic model scanning method

Aim Labs proposes a sandbox-execution-based dynamic model scanning method, which compensates for the inherent shortcomings of existing static scanners by executing models in a sandbox environment and tracking their operations. The method is a patent-pending technology and has been presented in the Black Hat USA 2025 talk "Smashing Model Scanners: Advanced Bypass Methods and a Novel Detection Approach". According to the same report, existing scanners rely on preset blacklists of functions and modules, making it difficult to cover millions of Python functions, and static code analysis is an NP-hard problem that is easily bypassed. Currently, public information about this method comes from this one talk, and no subsequent progress has been disclosed.

AI-generated from public reports

1. Aug 18· 1 reports

   Aim Labs proposes a sandbox-execution-based dynamic model scanning method

   Source:Cato AI Labs