Skip to content
Nubent

/Gartner · Apr 9, 2026

Gartner: MCP and agents drive AI security incidents

Gartner predicts that by 2028, 25% of enterprise generative AI applications will experience at least five minor security incidents per year, up from 9% in 2025.

What changes

According to Gartner, the jump comes from agents and MCP: the protocol was designed for fast, easy interoperability, not for security by default, so mistakes show up in normal use. By 2029 it expects 15% of those applications to have at least one major incident per year, up from 3% in 2025. The biggest risk sits in workflows where an agent accesses sensitive data, processes untrusted content and communicates externally at the same time: Gartner recommends treating that combination as a no-go zone.

How to implement it

  1. Inventory the agents and MCP servers already in use at your company, sanctioned or not.

  2. Set up a formal security review for each MCP use case and start with the low-risk ones.

  3. Give each agent its own identity and narrowly scoped permissions instead of inheriting a user's.

  4. Defend against prompt injection and review third-party MCP components like any other dependency.

Nubent's take

It matches the most common mistakes we see when teams connect agents: giving them a person's role, installing MCP servers like just another plugin, and letting a workflow read external email while it has access to customer data. The defense isn't to stop AI, but to limit what each agent can do and log every call.

Next step

Should we implement it in your account?

We check whether it fits, how much work it takes and what risks it carries. 30 minutes with an engineer.