Death by (Python) Pickle: "Betrayal ML"
summary
Claim: Python pickle files can carry trojans that backdoor models; examples include PyTorch supply-chain compromises in 2022 and 2023 and a pandas release (1.3.0 RC1) that shipped malware. ReversingLabs uses a searchable 24 petabyte repository and a free checker at secure.software to detect tampering and typosquatting (cryptog... vs cryptography) and recommends automated behavioral diff and dependency scanning for Hugging Face and pip models. These defenses scale and have found real backdoors in top‑100 packages.
AI-written summary · about
chapters (16)
- 0:00AI as attack surface
- 3:35Untrusted inputs problem
- 5:02What is ML model
- 7:55Impossible AI scenarios
- 12:34Pickle files matter
- 15:10Detecting AI output anomalies
- 16:59Supply chain security
- 19:31Roll your own crypto in AI
- 22:00Open source research
- 23:50Malware in PyPI packages
- 27:25Hardening defects risks
- 30:21AI-generated code security
- 31:24EU Cyber Resilience Act
- 33:51Supply chain testing
- 34:54Final quiz recap
- 37:23Q&A on CRA
related talks
32mPoisoned Pickles Make You Ill- 32mPoisoned Pickles Make You Ill
27mPoisoned Pickles Make You Ill
56mHidden Risks in Open-Source Code and AI Models
37mBeyond the Basics: Hacking and Securing Machine Learning Systems and Environments
36mThe Dark Side of AI: The Hidden Risks in Open-Source AI Models
50mMalice In Chains: Supply Chain Attacks Using Machine Learning Models
45mThe Dark Side of AI: The Hidden Supply Chain Risks in Open...