Google Revolutionizes Agent Security with Egress Proxy Model

Google has made significant advancements in agent security with its recent update to Managed Agents in the Gemini API. The September 2026 preview introduced a new Credentials API architecture that ensures sensitive secrets never reside within the ephemeral Linux sandbox where agents execute.

The egress proxy model utilized by Google performs HTTP header transforms, such as injecting an Authorization: Bearer token, on the fly, without exposing raw tokens to the sandbox code or environment variables. This approach represents a distinct strategy for mitigating risks inherent in autonomous agent execution.

Google's development is part of a broader trend in agent identity and governance stack formalization. In one week in September 2026, four major vendors introduced four distinct credential architectures, highlighting the industry's focus on securing agentic workflows. These moves signal that the industry is moving beyond experimentation and into rigorous infrastructure hardening.

The governance stack now consists of four functional layers: build-time policy enforcement, runtime management, data-security layers, and managed harness layer represented by Google. By providing a standardized environment for credential lifecycle handling, Google positions itself as a foundational provider in this emerging stack.