AI agents in companies: CEO of Bardoor AI warns of 'governance crisis' with malicious MCP servers

Bardoor AI CEO Oren Michels sounded the alarm in an interview with CNBC's Fortt Knox program: the new wave of corporate AI agents connected to corporate systems is generating a silent governance crisis. Thousands of MCP (Model Context Protocol) servers are already circulating in business environments, and some of them contain malware that can exploit excessive permissions. The solution proposed by Bardoor AI is a unique control panel that inspects LLM and MCP calls before they reach critical systems such as Salesforce, Gmail or internal databases — an approach Michels calls "firewall for AI agents".

AI agents: What are MCP servers and why do they generate risk?

MCP, or Model Context Protocol, is an emerging pattern that allows language models (LLMs) to access data and actions of external systems in a standardized manner. In practice, each integration with a CRM, ERP or corporate API may require an MCP server. The problem, according to Michels, is that many of these servers are created by third-party developers or even by employees without IT supervision, creating a huge attack surface. "We are seeing thousands of MCP servers being deployed without any security control," the executive warned. Lack of governance paves the way for malicious agents to inject dangerous commands — such as extracting customer data or modifying financial records.

How does Barndoor AI solve the governance crisis of AI agents?

The startup is developing a control panel that acts as an intermediary between LLM and corporate systems. The tool inspects each call — both model request and MCP server response — in real time by applying predefined security rules. "Our goal is to give full visibility to what each agent is doing and block unauthorized actions before they cause damage," Michels explained. The proposal is aligned with broader initiatives of Intelligence and Trust: the two pillars of the AI according to Microsoft, which highlight the need for transparency in autonomous systems.

AI agents: Why should Brazilian companies worry?

With the rapid adoption of AI agents in Brazil — especially in sectors such as fintechs, retail and health — the absence of governance can become a regulatory and operational liability. Estonia, for example, has already created an official identity for AI agents, defining rights and responsibilities by law. Without similar mechanisms, Brazilian companies run the risk of exposing data sensitive to unaudited MCP servers. Michels recommends that information security teams conduct complete inventories of all AI agents in operation and implement minimum-privileged policies — the same principle used for user accounts.

How to identify exposure to servers MCP malicious with AI agents?

Warning signals include: AI agents that have unrestricted access to critical systems, absence of call logs between model and APIs, and MCP servers with database reading and writing permissions. Barndoor AI suggests a simple audit: list all active MCP endpoints and check that each has strong authentication and documented access policy. Otherwise, the company may be vulnerable to attacks such as what Pramaana Labs tries to mitigate with formal AI verification.

Frequently Asked Questions about AI Agents and Governance

What is an MCP server?

It is a software component that implements the Model Context Protocol, allowing an LLM to access data and perform actions on external systems such as CRMs, emails or databases.

How can an MCP server contain malware?

Like any software, MCP servers can be created by malicious agents and hosted in public repositories. When deployed, they may execute unauthorized commands such as data exfiltration or modification of records.

What should companies do now?

Perform an inventory of all AI agents in production, map associated MCP servers, implement strong authentication and consider traffic inspection tools between LLM and systems — such as the panel proposed by Bardoor AI.

Source: [CNBC - Oren Michels, Barndoor AI CEO, Fortt Knox conversation] (https://www.cnbc.com/video/2026/06/17/oren-michels-barndoor-ai-ceo-a-fortt-knox-conversation.html)