← Back to security

Open-Source AI Model Poisoning & Backdoors

6 questions · advanced

Test your understanding of PyTorch pickle reduce deserialization exploits, Safetensors binary header layout, neural backdoor attention weight perturbations, SFT dataset poisoning mechanics, cryptographic SHA-256 commit verification, and container sandbox isolation for untrusted AI model weights.

0 / 6

How does arbitrary code execution occur during the deserialization of untrusted PyTorch .bin or .pt model checkpoint files stored in legacy Python pickle format?

Press 1 to 4 to pick an answer