Attestation
MCP Gets Enterprise Auth — Audit Logging Still Isn't Part of the Deal
The Model Context Protocol just picked up a real enterprise feature: centralized, identity-provider-backed authorization, the kind that lets a company route agent access through Okta or Entra ID instead of ad-hoc API keys. An IAPP analysis by Kapil Duraphe credits MCP for that upgrade while making a sharper point: controlling what an agent can reach is not the same as recording what it actually did.
The gap shows up at the moment an incident review starts. When an agent deletes the wrong records or pulls a table it shouldn't have touched, the first two questions are what did it do and who allowed it. MCP's access controls answer the second question better than before. They don't answer the first. A companion review of six major agent platforms — Microsoft Foundry, Anthropic's Claude products, OpenAI, Google's Gemini Enterprise, AWS Bedrock, and the MCP spec itself — found that MCP defines no audit log at all. It specifies a logging utility for servers to send diagnostic messages to clients and flags that token passthrough makes investigation harder, but leaves everything evidentiary to whoever implements the server.
That matters because MCP has become the default connector layer for enterprise agents precisely because it standardized access. Before it, each vendor wired agents into internal systems its own incompatible way. Removing that friction also removed an accidental security boundary: previously, an agent simply couldn't reach a system nobody had bothered to connect it to. MCP erases that constraint by design, and nothing has stepped in to replace it at the evidence layer.
AWS's Bedrock AgentCore is reportedly the only platform among the six whose native audit log documents individual tool calls by name — and even there, arguments passed to the tool are redacted by default, which is a privacy choice with real forensic cost. Everyone else leaves tool-call detail to customer-built telemetry.
The practical upshot for compliance teams: enterprise-managed authorization in MCP is a genuine step toward zero-trust agent access, but it's an access-control upgrade, not an accountability one. Teams relying on it to satisfy audit or incident-response obligations are building on a layer that was never designed to answer 'what did the agent do and why' — only 'was it allowed to connect.'