What MCP Apps actually change
The Model Context Protocol standardizes how AI applications connect to external tools, data sources, and services through a client-server architecture. Anthropic's MCP Apps extend this by allowing agents to invoke third-party tools interactively during a conversation, turning what was previously a developer-configured integration into a runtime decision made by the agent itself. This shifts the governance question from "which systems did we integrate" to "which systems can an agent decide to call, and under what conditions." For enterprises, that is a meaningful change in attack surface. A connector that was safe as a static integration becomes a live decision point every time the agent runs, and the set of tools available to it can change dynamically at runtime, which complicates any governance model built around fixed, pre-approved permission lists.
What the protocol does and does not guarantee
MCP defines an authorization framework based on OAuth 2.1 for HTTP-based transports, including mechanisms for authorization server discovery and protected resource metadata so that MCP clients can locate and authenticate against the correct authorization server. Token audience binding and resource indicators have been added to the specification to prevent a token issued for one MCP server from being replayed against another, a class of issue known as a confused-deputy problem. These are real technical mechanisms, and they matter. But the specification is explicit that implementing proper security controls is the responsibility of the MCP server and client implementers, not something the protocol itself guarantees. That distinction is the core of the enterprise governance gap: the protocol provides the vocabulary for authorization, but not the enforcement.
Where user-level approval falls short
Anthropic's Claude apps require explicit user approval before a tool call executes, with an option to enable auto-approval for specific tools. This is a reasonable default for individual use, but it does not function as an enterprise control. Per-user approval decisions are inconsistent across an organization, are not centrally auditable in most deployments, and can be effectively disabled at scale if auto-approval is turned on broadly for convenience. Anthropic also publishes a directory and review process for MCP connectors used within Claude, which addresses supply-chain vetting of published integrations, but it does not replace an enterprise's own policy about which of those connectors, once approved externally, should be allowed to touch internal data in a given environment.
Known risk categories specific to MCP
MCP documentation identifies several threat categories relevant to tool-calling agents: tool poisoning, where a malicious or compromised server misrepresents what a tool does; prompt injection delivered through tool names or descriptions rather than user input, which is a distinct vector from traditional prompt injection; and confused-deputy scenarios where a token or credential intended for one resource is misused against another. These risks are amplified in enterprise settings because MCP separates the client, meaning the agent or application, from the server, meaning the tool or data provider, and permission enforcement can occur at either layer depending on implementation. Without a clear decision about where enforcement lives, organizations can end up with no enforcement layer at all, relying instead on the good behavior of individual client applications.
Applying existing governance frameworks
No MCP-specific regulatory standard exists as of this writing. General-purpose frameworks such as NIST's AI Risk Management Framework identify access control, monitoring, and traceability of AI system actions as core functions applicable to agentic tool use, but they require organization-specific mapping to MCP's actual mechanics. In practice, this means treating MCP tool calls as first-class events in existing audit and compliance processes, equivalent in rigor to API calls or privileged user actions, rather than as a new category exempt from established logging and access review requirements. Governance leaders should map their existing Govern, Map, Measure, and Manage functions to specific MCP controls: who can grant a server access, how scope is defined, how token issuance is logged, and how revocation is executed.