CyberCertPrepAI Security FundamentalsAdversarial ML & Model AttacksEASYFree question
A hospital privacy officer is reviewing third-party model weights downloaded from a public repository. The weights are distributed as a .pkl file. What is the primary security concern?
- A.Pickle files lack version metadata, making model provenance impossible to verify
- B.Pickle files do not support model quantization, reducing inference accuracy
- C.Pickle files can execute arbitrary code during deserialization, allowing a malicious model distributor to run commands on the hospital's inference server
- D.Pickle files are incompatible with HIPAA-compliant encryption standards
Why C is correct
Python pickle files execute Python code during deserialization via the __reduce__ method. A malicious .pkl file can contain embedded commands that execute on the loading machine-a serious supply-chain threat. Downloading untrusted .pkl files and loading them with torch.load() or pickle.load() has led to real-world compromises. The other options describe inconveniences but not security threats; none relate to pickle's fundamental deserialization code-execution risk.
Know someone studying for AI Security Fundamentals? Send them this one.