Google ships managed agent harness with credentials never touching sandbox

Google’s managed‑agent preview (antigravity-preview-09-2026) separates credential handling via a proxy that injects auth headers at egress, preventing tokens from being exposed to agent sandboxes — a developer security and governance milestone.

In this brief: 2 sections 1 min read

- Credentials are handled by an egress proxy that performs HTTP header transforms (e.g., Authorization: Bearer) on the fly.
- Sensitive tokens never reside in the ephemeral Linux sandbox where the agent runs.
- Default model in the preview is upgraded from Gemini 3.5 Flash to Gemini 3.8 Flash.
- Configuration managed through version‑controllable AGENTS.md and SKILL.md files.
- Files can be mounted from Git repos or Google Cloud Storage to define persona and skill modules.
- Positioned as part of a formalizing governance stack for managed agent deployments.

Read full analysis on forkast.news

Useful?