Malicious ML models discovered on Hugging Face platform
Malicious ML models discovered on Hugging Face platform
www.reversinglabs.com
Malicious ML models discovered on Hugging Face platform
6 comments
-
-
Without reading the article, as a practicing AI Engineer here’s a couple of easy best practices:
- Use only the .safetensors files, the format is engineered specifically to allow only specific information necessary for NN frameworks
- Don’t use the ‘trust_remote_code=True’ parameter when serving your models without due consideration of the source of the model.
-
We've known pickle files have been unsafe for like three years at this point and people are still using them?
-
Three years? The last time I used pickle was for a school project over a decade ago and even then these vulnerabilities were clearly laid out in the documentation, and it strongly advised against using it for any serious application. The only reason I kept using it in the project is precisely because it was a school project, and I knew the application would never be used in any production context worth attacking. Watching the ML community enthusiastically embrace pickle in the time since has been very amusing to say the least. Honestly I'm surprised it only seems to be catching up to them now.
-
6 comments
This is a direct quote from the article:
In other words, there's a known vulnerability, it's documented, it's ignored and now it's been exploited twice.
Wow .. shocked .. is not a word I'd use to describe this situation.
Fuck around and see what happens .. seems more apt.
Twice that we know of.
Fair point.