Attestation
MCP Fixed Who Can Call a Tool. It Didn't Fix Who's Accountable After.
The Model Context Protocol just got a real upgrade to its access-control layer, and a new IAPP analysis published September 30 points out what it still doesn't do.
MCP's most recent release added enterprise-managed authorization: a centralized control plane that lets organizations gate agent tool access through their existing identity providers, Okta and Entra ID among them. That closes a gap that's dogged MCP deployments for over a year — tools and servers could be invoked with no consistent way to tie that invocation back to an enterprise identity policy. Now an admin can say which agents are allowed to call which tools, enforced at the IdP layer rather than left to whatever the MCP server implements on its own.
What the upgrade doesn't touch is what happens after the call is authorized. MCP has no built-in standard for tracing a tool invocation through to its outcome, or for logging what an agent actually did with the access it was granted, in a form that holds up for compliance or incident review. Authorization answers "was this agent allowed to call this tool." It doesn't answer "what did it do once it got there," or give auditors a consistent artifact to check against when something goes wrong.
That distinction matters more than it sounds. Access control and audit logging solve different problems and have historically been separate layers in enterprise IT — think RBAC versus SIEM. MCP standardized the first one. The second is still left to whatever logging each server or platform bolts on independently, which means two enterprises running the same MCP-based agent stack can end up with wildly different, and possibly incompatible, audit trails for the exact same category of action.
For teams treating agent deployments as something regulators or internal auditors will eventually ask about, this is the gap that doesn't close itself. An IdP-enforced permission model is necessary infrastructure, but it produces a yes/no gate, not a record. Without a logging and tracing standard that travels with the protocol the way authorization now does, every enterprise is left reconstructing accountability after the fact — from whatever their particular MCP server happened to write down, if anything.
The practical result: the protocol that's becoming the default way agents reach tools now has strong authorization and a self-assembled, inconsistent accountability layer sitting underneath it.