Theory · 8 min
The AI supply chain
Treat models, datasets and MCP servers as software you install, with the same risks.
Models, datasets, plugins and MCP servers are downloaded from public hubs and package registries, just like libraries. OWASP lists this as LLM03 Supply Chain; ATLAS as AML.T0010 AI Supply Chain Compromise.
Some model file formats can run code. Python’s pickle format, used by older PyTorch checkpoints, can execute arbitrary code when loaded; Hugging Face scans for this and recommends the safetensors format, which stores only tensors.
Poisoned training or fine-tuning data (LLM04, ATLAS AML.T0020) can plant behaviour that only appears on a trigger, which ordinary testing does not find.
MCP servers are often installed with one npx or pip command that runs whatever version is current. Pin versions, prefer verified publishers, and watch what the server process does, as in the case on this site where an agent ran npx -y on an unknown MCP package.
Your turn: decide on a scenario. A data scientist downloads model.pkl from an unknown account on a model hub and loads it with pickle. What is the main risk?
Do the interactive check Free, in the browser.
Sources
- Pickle scanning (why pickle files can run code) Hugging Face
- OWASP Top 10 for LLM Applications 2025 OWASP GenAI Security Project
- ATLAS data (technique IDs and names) MITRE
- ATT&CK T1195.002: Supply Chain Compromise: Compromise Software Supply Chain MITRE ATT&CK v20