What happened
In a November 20, 2026 announcement, Baseten engineers Stephen Day and Hao Wu explain how the company applies a common security standard across a fleet of more than 100 clusters spanning over 20 clouds, including more than a dozen neoclouds. Baseten says its Multi-cloud Capacity Management system treats GPUs as fungible across providers, allowing workloads to move among clusters. Because providers hand over infrastructure with different levels of hardening, the company evaluates each cluster and adds controls before it serves traffic.
Baseten frames its threat model around four trust boundaries: the infrastructure provider, operating partners, customer workloads, and internal access. At handover, it closes public ports except those required to serve traffic, places the Kubernetes API behind an allowlist, and requires key-only SSH with regularly rotated keys. SSH access is on demand, requires multi-factor authentication, and uses no long-lived credentials. A baseline scan must pass before a cluster joins the fleet. In operation, cluster traffic uses secure channels, node joins are protected against machines adding themselves, and hosts and Kubernetes are patched on a set cadence. Baseten says it wipes customer data before machines return to a provider.
Controls for third-party operators follow least privilege: providers and partners receive only the access needed for their layer, provider consoles require hardware-bound MFA, and partners notify Baseten of control-plane changes. Customer workloads are treated as untrusted. Each customer build runs in a microVM with its own kernel; customer and platform images use separate registries and pass through Baseten’s build platform. Network policy separates tenants, cross-cluster traffic uses mutual TLS, and a node agent authorizes direct RDMA transfers between nodes. Customer secrets are encrypted in Baseten’s control plane, segregated by organization, and delivered only to that customer’s deployments. Runtime detection runs on every node, including for GPU-runtime breakout attempts.
The company says credentials are scoped to at most one cluster to contain a compromise. Each cluster has its own certificate authority; its configuration-pull key is read-only, and distribution credentials rotate within an hour. Kubernetes audit, host, DNS and SSH logs from each cluster flow to a central monitoring system with alerts for an on-call responder. Baseten also conducts daily external perimeter scans and regularly checks provider security posture. Its checklist covers credential scope, infrastructure, partners and workload isolation, for use in auditing a fleet or assessing a provider.
What the archive adds
New Runtime’s earlier coverage of Tamnoon’s AI-powered cloud security remediation focused on a remediation pipeline as the operative unit of cloud-security automation. Baseten’s account is a new event: it adds a fleet-wide infrastructure-security practice, rather than another remediation workflow.
Trend signal
This supports agent security moves to runtime boundaries: Baseten describes security as enforceable boundaries around what workloads can access and do, backed by isolation, scoped credentials, logging and monitoring—not just instructions to behave safely. The account concerns inference infrastructure and customer workloads, rather than a general agent runtime, so the connection is relevant but not a complete match.
Why now
Baseten says its multi-cloud capacity model lets it use providers with different operating models and security defaults, including neoclouds. The need to establish a consistent baseline follows from that architecture: controls must hold wherever a workload lands, not only on infrastructure with uniform managed defaults.
Why it matters
For teams running AI workloads across clouds, provider-level assurance is not a substitute for workload-level controls. A cluster can arrive exposed, a partner may operate part of its stack, and customer code can be hostile or compromised. Baseten’s design makes the failure case explicit: one cluster may be taken over, but its credentials should not unlock another cluster or the control plane.
The transferable idea is to make the unit of trust small and observable. Separate keys and certificate authorities, read-only configuration access, isolated builds, controlled secret delivery and centralized logs turn broad security promises into checks that can be applied cluster by cluster. Teams can use the checklist to find gaps in their own fleet or ask an infrastructure provider precise questions before trusting it with models and data.
Limits and uncertainty
These controls are described by Baseten as its own practices; the account does not independently verify their implementation or effectiveness. It provides no measured incident reduction, audit results, or comparison of security outcomes among providers. The company names several control types and operational checks, but readers would need to validate how those controls are configured and tested in the environments they use. Requirements such as provider-console MFA, separate trust roots and hourly credential rotation may also depend on provider capabilities and workload architecture.
Skill to build now
Audit one representative cluster this week against Baseten’s checklist. Confirm that its public ports are limited to traffic-serving needs, its Kubernetes API is allowlisted, and its configuration-pull credential is read-only and unique to that cluster. Then trace how customer secrets reach workloads and whether logs leave the cluster. Record any control that cannot be verified, and ask the provider or platform owner for the concrete check that would prove it.
Reader takeaway
A multi-cloud AI fleet is only as consistent as its least-hardened cluster. Give each cluster narrowly scoped credentials, isolate workloads and verify controls continuously; do not assume that a provider’s defaults make clusters equivalent.