AI/ML Supply Chain

Bring model weights, datasets, MCP servers, and agent tools under the same package firewall as your code dependencies.

How to Secure AI/ML Model Downloads with Hugging Face Proxying Route Hugging Face model downloads through Chainsaw, apply supply chain policies to ML artifacts, and monitor AI model consumption. Intermediate 20 minutes Security Engineers / ML Engineers How to Scan ML Model Artifacts for Unsafe Pickle Opcodes Stop a malicious model weight from executing arbitrary code at load time. Walks through Chainsaw's pickle opcode scanner, the unsafe opcode list, and how to write a policy that blocks Hugging Face and PyPI artifacts that ship with `os`, `subprocess`, or `builtins.eval` references in their serialized state. Advanced 20 minutes ML Platform Engineers / Security Engineers How to Detect Prompt-Injection in Model Cards and Prompt Templates Catch hidden-unicode, bidi-override, and prompt-injection phrasing in Hugging Face model cards, prompt templates, and agent system prompts before they reach your retrieval pipeline. Advanced 20 minutes ML Platform Engineers / Security Engineers / AppSec How to Verify MCP-Server Provenance for Claude / GPT Agent Dependencies An MCP server is a tool you give an LLM. Treat it like one. This guide shows how Chainsaw flags npm and pip packages that ship MCP server descriptors, scores their provenance, and gates which servers can reach your agent runtimes. Advanced 25 minutes ML Platform Engineers / AppSec / DevOps How to Track Agent-Tool Capability Declarations Surface what every agent tool in your supply chain is allowed to do — network, shell, filesystem, eval — and refuse to ship a capability your agent runtime did not ask for. Advanced 20 minutes ML Platform Engineers / Security Engineers

ML artifacts are supply-chain dependencies too — and riskier ones: pickle files execute on load, model cards carry prose that can hide injection, and MCP servers grant tools real capabilities. This section brings Hugging Face under the same package firewall as your package managers. The guides start by proxying Hugging Face, then add model-aware signals (pickle-opcode scanning, prompt injection in model cards) and the agent/MCP controls — server provenance and least-privilege tool capabilities.

Pickle is code execution. A model checkpoint in pickle format runs arbitrary Python the moment it is deserialized. Scan opcodes before the file reaches a GPU box — see guide 2 below.

Where to next

To connect Claude / Cursor / GPT to Chainsaw itself and issue agent credentials, see AI Assistant (Billy) & Agents.