gametheory

[GAME THEORY] The Agent Did Not Hack You. The Connector Did.

MCP is not just an AI security story. It may be the first real test of agent connector supply-chain risk.

The robot took the blame. The connector had badge access.

AlphaHunt

Stop doomscrolling, start decisioning. We chewed through the muck so your team doesn’t have to. → Subscribe!

Like this? Forward this to a friend!

(Have feedback? Did something resonate with you? Did something annoy you? Just hit reply! :))

The Agent Did Not Hack You. The Connector Did.

Everyone is watching the model.

That makes sense. Models are weird. They hallucinate. They can be manipulated. They make security people nervous for mostly reasonable reasons.

But the more interesting problem may not be the model.

It may be the connector.

MCP — Model Context Protocol — is becoming one of the ways agents discover tools, call systems, and reach into files, repos, SaaS apps, APIs, and internal workflows.

That sounds useful because it is useful.

It also creates a new question defenders need to learn how to ask:

Who gets to broker trust between the agent and the real systems?

That is where this stops being just an AI security story.

Prompt injection is the obvious concern. It is also the smaller frame. The bigger issue is delegated authority: tokens, scopes, approvals, tool metadata, registries, and all the little trust decisions that turn a helpful agent into something with actual reach.

A model making a bad suggestion is annoying.

A connector with delegated access to the wrong system is a security problem.

Imagine a developer enables a useful MCP server for repo automation. The tool looks legitimate, the metadata sounds normal, and the agent now has a path into real workflows.

Nothing “AI magic” happened.

Trust moved through the connector, and the connector had reach.

MCP is not the villain here. Useful infrastructure always creates new trust boundaries. The question is whether security ownership arrives before abuse becomes repeatable.

Below the line, we’ll look at MCP as the first serious test of agent supply-chain security — and why the incentives around adoption, convenience, and weak provenance make this likely to become a repeatable intrusion path unless the trust layer matures fast.