Google publishes a three-layer zero-trust agent pattern

Google published a zero-trust reference pattern for agents with three layers outside the model: hardware-backed signatures for database writes, gVisor isolation for generated code, and deterministic gateways for inputs and outputs. The Google Developers Blog example caps sandboxed code at 64 MB, 0.1 CPU, and a 5-second timeout, illustrating the kind of hard boundary a prompt cannot provide. Builders moving agents toward refunds, production data, or code execution should file the pattern as an architecture checklist—not a guarantee that heuristics catch every attack—and compare it with the archive’s Snowflake control-plane analysis.