mirror of
https://github.com/netbirdio/gvisor.git
synced 2026-05-22 17:12:49 -07:00
Edit Stable Diffusion blog post to mention the Hugging Face pickle exploit.
https://www.wiz.io/blog/wiz-and-hugging-face-address-risks-to-ai-infrastructure PiperOrigin-RevId: 622332942
This commit is contained in:
committed by
gVisor bot
parent
336193e33a
commit
6b93f10457
@@ -52,8 +52,10 @@ Diffusion) are packaged using the more secure [safetensors] format, **the
|
||||
majority of models available online today are distributed using the
|
||||
[Pickle format], which can execute arbitrary Python code** upon deserialization.
|
||||
As such, even when using trustworthy software, using Pickle-formatted models may
|
||||
still be risky. gVisor provides a layer of protection around this process which
|
||||
helps protect the host machine.
|
||||
still be risky (**Edited 2024-04-04:
|
||||
[this exact vulnerability vector was found in Hugging Face's Inference API](https://www.wiz.io/blog/wiz-and-hugging-face-address-risks-to-ai-infrastructure)**).
|
||||
gVisor provides a layer of protection around this process which helps protect
|
||||
the host machine.
|
||||
|
||||
Third, **machine learning applications are typically not I/O heavy**, which
|
||||
means they tend not to experience a significant performance overhead. The
|
||||
|
||||
Reference in New Issue
Block a user