Google's Agent Sandboxes Are the Right Answer to Prompt Injection

Google unveiled Agent Compute Engine, hardware-isolated sandboxes that never hold credentials. It is the most sensible response yet to prompt injection, the attack that has been quietly haunting agent builds.

Prompt injection is not a bug you patch once. It is a class of problem: a model reads untrusted input and acts on it. The only durable fix is to make sure acting cannot reach anything valuable.

A sandbox with no credentials to steal turns a catastrophic attack into a contained one. That is the right shape, even if it is not the final shape.

I would judge every agent platform by this question now: when it gets injected, what can the attacker actually reach? Google's answer is 'nothing.' That should be the floor, not the ceiling.