AI agents now have a software supply-chain problem
September 20, 2026 · Keith Pasley
Most of the AI security conversation has focused on the model: prompt injection, data leakage, guardrails and model behavior.
But agents are creating another problem.
They don’t operate alone. They connect to plugins, tools, APIs, MCP servers and other systems. Every one of those connections expands what the agent trusts.
The recently disclosed Plugin4Shell vulnerability is a good example. Researchers found weaknesses affecting several major AI coding agents that could allow malicious plugin code to get past controls intended to ensure the agent was using a trusted version.
I think this is where enterprise AI security gets more complicated.
It isn’t enough to secure the model or even the agent. We also have to know what the agent is connected to, what it trusts and what those connections allow it to do.
We’ve dealt with software supply-chain security for years.
Now we’re getting an agent supply chain.
And as agents get more access and autonomy, securing that chain is going to matter.
Technology doesn’t create business value. Better decisions do.