Google Cloud launches Agent Compute Engine: every AI agent gets its own empty computer
AspireAI · Google Cloud Blog
Google Cloud launched Agent Compute Engine (ACE) at AGNTCon, a purpose-built compute service already powering Gemini as an agent. Every agent gets its own hardware-isolated microVM with its own filesystem, language runtimes and browser, booting fast enough to sit inside an agent loop.
The security design is the notable part. Google's stated rule: assume an attacker controls everything inside the sandbox, then make sure they find nothing there to use. Guest environments hold no API keys, no OAuth refresh tokens and no cloud service-account credentials, and sandboxes boot without a standard network card, with all traffic mediated through a proxy. Long-lived user credentials stay off-host and are swapped in only after the session's identity is validated. Google also separates the agent harness (the brain that calls the model) from the sandbox (the hands that execute), so the model never runs inside the untrusted environment.
The launch also extends Gemini for business agent governance, including Agent Registry, per-agent identities and policy-driven outbound brokering.
Why it matters: the hard part of agent security is not prompt hardening. It is making sure that a successful prompt injection leaves the attacker with nothing to steal. ACE is built on that assumption rather than trying to prevent the injection.
Caveat: Agent Compute Engine is targeted for preview in the coming months. It is not generally available today.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.