60,000 sandboxed agents. One of them breaks out.
A guest-to-host escape through the vGPU path puts its neighbours and the management plane within reach.
When an AI workload tries to escape its VM, Claura’s defender agents seal it off like a ship’s bulkhead.
A guest-to-host escape through the vGPU path puts its neighbours and the management plane within reach.
Playbooks only match attacks they have already seen. This one is new.
Every minute spent waking someone, finding the host and deciding what to cut is a minute the breach spends spreading.
Each agent has one job. The one high-impact step stays with a person.
It correlates IOMMU faults, hypervisor audit logs, host EDR, IPMI flows and token misuse into a single incident.
One suspect host, 11 co-resident guests and a reachable management plane. Triage scores what the escape could touch before anything moves.
Cut the network, revoke credentials, cordon the host. Then pause the guest, capture its memory and fail the guests over to clean hosts.
Reversible first
Then
Quarantine host waits for a person
A forensic record of every signal, action and approval, kept while the others work.
Quarantining the host is the single high-impact step. Evidence, scope and rollback sit on one card. Approve with ⌘↵. Tap to approve.
Proposed by Warden · policy-bounded
Warden: qemu on this host spawned a shell after a DMA fault. Treat the host as untrusted until it is reimaged.
Warden cannot quarantine a host on its own.
Contained
0 co-resident guests affected. Six automated steps, one approval, one sealed host.
Defender agents reason with NVIDIA Nemotron, self-hosted on-prem, so telemetry stays in the building. Every proposed action passes a policy gate before anything runs.
Match attacks they have already seen, then run fixed steps.
Assemble evidence on attacks nobody has seen and propose bounded actions. Policy, not the model, decides what executes.