INTERNET STORM CENTER SPOTLIGHT
ISC provides a free analysis and warning service to thousands of Internet users and organizations, and is actively working with Internet Service Providers to fight back against the most malicious attackers. https://isc.sans.edu/about.html
ModelScan - Protection Against Model Serialization Attacks
Published: 2025-02-17
Last Updated: 2025-02-18 00:37:10 UTC
by Russ McRee (Version: 1)
Protect AI’s OSS portfolio includes tools aimed at improving security of AI/ML software. These tools are meant for a wide range of engineering, security and ML practitioners including developers, security engineers/researchers, ML engineers, LLM engineers and prompt engineers, and data scientists.
Of particular interest in light of model serialization attacks is ModelScan.
Headlines as recent as 6 FEB 2025 remind us that the popular Python Pickle serialization format, common for distributing AI models, offers attackers opportunities to inject malicious code to be executed when loading models with PyTorch. See Malicious ML models discovered on Hugging Face platform. Post training, model’s mathematical representations can be stored in a variety of data serialization formats to be shared and reused without the need for additional model training. Pickle is a popular Python module used for serializing and deserializing ML model data. While easy to use, Pickle is considered an unsafe data format, as it allows Python code to be executed during ML model deserialization.
As you can imagine, even as protective measures are being implemented, safety scanning is still recommended. ModelScan offers such capabilities with ease and convenience. ModelScan is incredibly well documented and include notebooks to aid experimentation and adoption.
I’ll share my quick setup steps, modify to your liking and preferences. These assume you’re building from scratch including Jupyter ...
Read the full entry: https://isc.sans.edu/diary/ModelScan+Protection+Against+Model+Serialization+Attacks/31692/
XWorm Cocktail: A Mix of PE data with PowerShell Code
Published: 2025-02-19
Last Updated: 2025-02-19 07:39:49 UTC
by Xavier Mertens (Version: 1)
While hunting, I spent some time trying to deobfuscate a malicious file discovered on VT. It triggered my PowerShell rule. At the end, I found two files that look close together ...
They are identified as “data files,” and their upload names are, respectively, “XClient<.>exe” and “XingCode Unblocker 2025<.>exe". XignCode is anti-cheat software primarily used in online games to prevent cheating, hacking, and the use of unauthorized third-party tools. Note the typo in the file name!
When you open the file, you see this:
Read the full entry: https://isc.sans.edu/diary/XWorm+Cocktail+A+Mix+of+PE+data+with+PowerShell+Code/31700/