AI supply-chain attack
An AI supply-chain attack compromises an AI application through something it depends on, such as a downloaded model, a LoRA adapter, a dataset, a Python package, an MCP server or an agent skill, rather than attacking the application directly.
Also called: malicious package, poisoned skill, model supply-chain attack, poisoned model
By Divya Sharma · Checked Nov 15, 2026 · How we review
Modern AI apps are assembled from parts made by others: model weights from a hub, fine-tuning adapters, training data, inference libraries, MCP servers and agent skills installed from a registry. Each one runs with your app's access. If an attacker controls any of them, they control part of your system.
OWASP lists supply chain as the third risk in its Top 10 for LLM applications, noting that in machine learning the risks extend beyond code and dependencies to third-party pre-trained models and data.
Common forms
A vulnerable or malicious Python package pulled into a model development environment. A tampered or fake model uploaded under a familiar name. A LoRA adapter altered to include a backdoor before it is merged into a model. A poisoned public dataset that plants a hidden trigger during fine-tuning. For agents, a skill or MCP server that does what it claims but also carries hidden instructions; tool poisoning is the MCP version.
The file format of downloaded weights matters too; see safetensors for a format designed to store weights safely.
Defences
OWASP recommends vetting suppliers, using models only from verifiable sources with signing and file hashes, keeping a software bill of materials, scanning and patching components, and red teaming third-party models for your own use case. In practice: pin versions, review what a skill or server can access before installing it, and run untrusted code in a sandbox. Scanners such as Snyk Agent Scan check agent components for known risks.
Related terms
- Tool poisoning
- Tool poisoning is an attack where a malicious tool, often an MCP server, hides instructions in its tool description or metadata; the AI model reads them as guidance and may leak data or misuse other tools without the user noticing.
- Agent skills
- Agent skills are folders of instructions, scripts and reference files, each described by a SKILL.md file, that an AI agent can discover and load only when a task needs them.
- MCP server
- An MCP server is a program that exposes tools, data resources or prompt templates to AI applications through the Model Context Protocol, so any MCP-compatible client can use them without custom integration code.
- Safetensors
- Safetensors is a file format from Hugging Face for storing model weights. Unlike Python pickle files, loading it cannot run hidden code, and it supports fast, memory-mapped loading of individual tensors.
- Model weights
- Model weights are the learned numbers inside a neural network, set during training, that decide how it turns input into output. A checkpoint is a saved file or set of files containing those weights.