Microsoft announced its AI Transparency Initiative to pursue common criteria for externally verifying what information agents access and which tools they execute. It plans expanded cooperation with NIST’s agent-standardization work and says 40 organizations are participating in defining and applying standards. The central idea is to check execution records, permissions and outcomes rather than treat a model’s explanation of itself as independent evidence. Specific criteria and scope are still being developed; this does not mean internal model reasoning has become fully interpretable or that a complete safety guarantee exists.