Skip to content
OWASP LLM Top 10 guide

LLM03:2025

Supply Chain

Risks from compromised or unreviewed models, datasets, plugins, packages, MCP servers, prompts, infrastructure, and vendors in an AI application.

Step 01

Input

Step 02

Model

Step 03

Tool / Data

Step 04

Impact

What it is

The AI system depends on components whose provenance, permissions, behavior, or update path is not controlled well enough for the application's risk level.

Why it matters

A compromised package or connector can use the application's access to steal data or run code. A tampered model or dataset can change answers while passing ordinary functional checks.

Failure path

One way this failure can develop.

Use this example to locate the input, permission or application check that would interrupt the sequence in your system.

Path 01

Compromise a dependency, model artifact, connector, dataset, package, or prompt template.

Path 02

Wait for the application to ingest or execute that component inside trusted workflows.

Path 03

Abuse inherited trust to change outputs, access data, or trigger downstream actions.

Defenses

Controls worth checking.

Check where each control runs, who owns it and how you will test it again after a prompt, model or workflow change.

Control 01

Inventory AI components

Track model providers, model versions, datasets, embedding models, MCP servers, plugins, packages, prompts, and runtime services.

Control 02

Verify provenance and updates

Pin versions, check signatures or hashes where practical, review release channels, and separate test upgrades from production rollout.

Control 03

Limit inherited authority

Do not let a third-party model, connector, or plugin inherit broad data or tool access by default.

Signals and control gaps

  • Unpinned model, package, or connector versions in production.
  • MCP tools or plugins added without owner, scope, and rollback path.
  • Model or embedding changes without regression results.

Questions for your team

  • What model and connector versions are currently trusted?
  • Who approves prompt, plugin, MCP, model, or dataset changes?
  • Can a dependency update silently expand data or tool access?