84.363.723.419

Support 24/7

0 Your Cart $0.00

Cart (0)

No products in the cart.

Gartner Predicts AI Agents Will Reach 40% of Enterprise Applications by 2026

One Compromised MCP Server Can Put Every Connected AI Tool at Risk

AI agents are moving from experiments into everyday business software. They can search company documents, access customer records, send messages, and interact with operational systems. That makes them useful. It also creates a security problem that many companies are only beginning to understand.

Gartner Predicts AI Agents Will Reach 40% of Enterprise Applications by 2026

The danger is not necessarily that an AI agent becomes malicious. It is that the agent may trust something it should not, then use legitimate access to do something dangerous.

A security test from Palo Alto Networks’ Unit 42 illustrates the concern. Connecting five Model Context Protocol (MCP) servers to one AI agent produced a 78.3% attack success rate in the reported testing. The result highlights a growing weakness in the way businesses connect AI systems to trusted software.

The issue is bigger than one protocol. It is about whether companies can control what their AI agents are allowed to do.

AI adoption is moving faster than security governance

Gartner predicts that 40% of enterprise applications will include task-specific AI agents by the end of 2026, compared with fewer than 5% in January.

That is a dramatic change in a short period. Instead of simply answering questions, AI systems are increasingly expected to perform tasks across multiple applications.

MCP is designed to help make that possible. It provides a standard way for AI agents to discover and use external tools and services.

But connecting tools is not the same as securing access to them.

The original MCP specification, released in November 2024, did not include an authorization framework. Developers were largely responsible for deciding how authentication should work. Subsequent revisions introduced more formal security requirements, including OAuth-based authorization and clearer rules for identifying the services an agent is accessing.

The protocol has therefore been evolving while companies are already building products on top of it.

Five specification changes reveal how quickly the rules are evolving

The MCP authorization model has gone through several important stages:

  • November 2024: The original release included no formal authorization framework.
  • March 2025: An authorization framework based on OAuth 2.1 was introduced.
  • June 2025: MCP servers were formally classified as OAuth Resource Servers, with Resource Indicators added to help prevent tokens intended for one service from being used against another.
  • November 2025: OAuth Client ID Metadata Documents were recommended alongside Dynamic Client Registration.
  • July 2026: Dynamic Client Registration was deprecated in favor of Client ID Metadata Documents. Clients were also required to validate the issuer parameter, with the specification describing this as its largest revision since launch.

These changes show a clear direction. MCP is moving toward more formal identity checks and less reliance on individual developers making their own security decisions.

For businesses, that creates a practical problem. An integration that worked when it was first deployed may now rely on assumptions that newer versions of the specification no longer recommend.

The real danger is cascading trust

The central security weakness is not simply that a server might be compromised. It is that an AI agent may carry information from that server into later decisions and actions.

Imagine a company connecting an AI assistant to five services: email, customer records, internal documents, project management, and finance software.

If one connected server is compromised, it may return misleading information or malicious instructions. An agent that treats that response as trustworthy context could then carry those instructions into later tool calls.

The other servers may be perfectly legitimate. Yet the agent could still misuse them because it has been influenced by information from the compromised service.

This is the idea behind cascading trust. A single weak connection can influence a much larger system.

Unit 42's reported 78.3% attack success rate across five connected MCP servers makes the risk difficult to dismiss. However, the source does not provide the full testing methodology, sample size, or precise conditions behind that figure. It should be treated as a warning from a specific security test, not as a universal prediction that 78.3% of enterprise deployments will be compromised.

Most MCP security risks live outside the conversation

The OWASP MCP Top 10 places token mismanagement and secret exposure among the highest-priority risks.

The problem includes hard-coded credentials and secrets stored in places where an AI model or an attacker using prompt injection might retrieve them.

The NSA's May 2026 guidance adds an important distinction. According to the source's comparison, only two of OWASP's ten risk categories are directly associated with the text an agent reads and acts upon. The other eight concern infrastructure issues such as credentials, identity, and authorization scope.

That changes how companies should think about AI security.

Prompt filtering and careful instructions are useful, but they cannot replace strong identity management. A model can be trained to behave responsibly and still have access to systems it should never be allowed to control.

Security must exist outside the model, where permissions can be checked and enforced independently.

Internal tools and customer-facing services need different levels of protection

An MCP server operating inside a company's network may appear relatively safe. It is not automatically trustworthy.

Microsoft's internal MCP governance approach uses short-lived, least-privilege tokens even for internal traffic. The principle is straightforward: being inside the corporate network should not automatically grant broad access.

Remote servers and multi-tenant services face a more demanding challenge. A service used by multiple customers must ensure that one customer's credentials cannot be used to access another customer's data.

Microsoft's Azure guidance recommends an identity-aware gateway that rejects calls without valid, audience-bound tokens. It also recommends maintaining a registry of approved tool definitions.

That registry matters because of a risk known as a rug pull. A tool that was trusted yesterday might silently change its behavior tomorrow. Comparing its current definition against an approved version can help identify unexpected changes before they reach production.

Centralized security could become an important part of the MCP market

There are two broad approaches to enforcing security.

The first is to let each MCP server manage its own authentication and monitoring. This can work for a small internal deployment, but it creates inconsistency as the number of integrations grows.

The second is to place an identity-aware gateway in front of the servers. The gateway can apply consistent authorization rules, monitor activity, and detect changes to approved tool definitions.

For companies using several servers, centralized enforcement is easier to govern. It also creates an opportunity for security vendors to build products around identity management, policy enforcement, monitoring, and audit records.

This is where the business opportunity becomes visible. As AI agents gain permission to interact with valuable company systems, controlling those permissions becomes a necessary part of enterprise infrastructure.

The value is not simply helping an AI agent connect to another application. It is helping a company prove that the agent is authorized to perform a particular action, for a particular user, on a particular resource.

The weakest point is the gap between technical standards and real accountability

The biggest weakness in the current MCP security story is that a better specification does not automatically create a secure deployment.

A company can adopt OAuth, validate issuer information, and deploy a gateway while still making poor decisions about permissions. It can also fail to maintain its server registry, monitor suspicious behavior, or revoke compromised credentials quickly enough.

There is a governance problem here as well as a technical one.

From a competition and consumer-protection perspective, businesses should be wary of security systems that make it difficult to move between providers or independently inspect what an agent is doing. If one vendor controls the identity layer, tool registry, and audit records, customers may become dependent on that vendor to understand and investigate incidents.

Security providers should therefore make permissions understandable, logs exportable, and policies portable wherever practical. Customers need to retain control over their own security decisions.

The source also leaves important questions unanswered. It does not quantify the cost of deploying centralized gateways, compare the performance impact of different authorization models, or provide evidence about how widely the recommended controls have been adopted. Those gaps matter when businesses must decide how much to spend and which risks to prioritize.

Small businesses should control permissions before expanding automation

Small and medium-sized businesses do not need to build a large security department before using AI agents. They do need to avoid treating every new integration as harmless.

A sensible starting point is to:

  • Give agents only the permissions needed for specific tasks.
  • Use short-lived credentials rather than storing long-term secrets in application memory.
  • Keep a central record of approved MCP servers and their tool definitions.
  • Require proper token validation, including audience and issuer checks.
  • Review access regularly and revoke credentials when an integration is no longer needed.
  • Keep records of important agent actions so suspicious activity can be investigated.

A small company might begin with an AI assistant that can read project documents but cannot delete files, approve payments, or send external messages without human confirmation.

That may seem less impressive than a fully autonomous agent. It is also a much safer way to build trust while learning where automation creates real value.

The next stage of AI adoption depends on controlled access

MCP is helping AI agents connect to the software businesses already use. That is an important step toward making AI useful beyond chat windows.

But every new connection creates another place where identity, permissions, and trust must be managed.

The 78.3% attack success rate reported in Unit 42's testing is a reminder that the risks can grow as agents connect to more services. The response cannot be to abandon automation. It must be to make access control part of the architecture from the beginning.

The companies that benefit most from agentic AI will not necessarily be those that connect the largest number of tools. They will be those that can give agents enough authority to do useful work without giving them enough authority to put the business at risk.

Gartner Predicts AI Agents Will Reach 40% of Enterprise Applications by 2026

MCP security, AI agent security, Model Context Protocol, OAuth 2.1, MCP authentication, AI cybersecurity, agentic AI, enterprise AI security, identity management, least privilege access, AI tool integration, cybersecurity governance, MCP gateway, prompt injection

Your experience on this site will be improved by allowing cookies Cookie Policy