Your guests run on a provider's virtualisation cluster whose virtual switch you cannot filter at — where do you enforce against a guest-to-guest pivot, and what does it cost?

## The constraint that makes this a real question On a managed provider's virtualisation cluster you rent guests. The provider owns the hypervisor, the virtual switch and the physical fabric. The layer with the best position for guest-to-guest enforcement is therefore not yours, and no amount of design inside the guest can move it. The interviewer wants to see you enumerate the remaining positions honestly, rank them against a specific intrusion, and quote a price for each rather than reaching for a favourite answer. ## The intrusion to rank against Be concrete: an intruder holds local administrative privilege in one guest and wants the neighbour — a database, a build host, a management VM — that happens to be scheduled on the same hypervisor. The flow will be switched in software inside the host. Nothing you operate on the physical network is in its path. ## Option 1 — mutual deny-by-default agents Put a firewall agent in every guest and write policy from both directions: the sender may only reach declared destinations, and the receiver only accepts declared sources. - **Buys:** process- and user-level policy; works regardless of where the provider schedules the guest; nothing to negotiate with the provider. - **Costs:** a fleet to install, keep running and keep in policy; CPU inside every guest; and the honest limitation that the agent in the *compromised* guest is administered by the intruder. The enforcement you can actually count on is the receiver's, so the policy has to be written so that the far end alone is sufficient. ## Option 2 — ask the provider for a per-guest filter Many providers expose a per-guest filter enforced in their virtual switch. - **Buys:** an enforcement point outside the guest's operating system, so an intruder with root in the guest cannot disable it; it covers same-host guest-to-guest traffic, which is exactly the gap. - **Costs:** it enforces on addresses and ports, never on a process or a user, so "the reporting service may reach the database" degrades to "this address may reach that address on this port". Change velocity becomes theirs: a rule change is a ticket or an API call against their control plane, with their availability and their maintenance windows. You also inherit their failure behaviour, which you did not choose. ## Option 3 — move the traffic onto a path you own If the flow must cross a boundary you operate, all your existing controls apply again. Split the workloads across hosts with anti-affinity, into separate segments, or into separate tenancies or accounts. - **Buys:** the pivot now traverses a firewall you configure, log and can prove denies. - **Costs:** capacity is fragmented and less efficiently packed; cross-host traffic consumes fabric bandwidth and adds latency to a path that used to be a memory copy; and placement rules are a standing operational obligation — the day someone relaxes anti-affinity to fit a maintenance window, the assumption silently dies. ## Option 4 — compensate rather than filter Where reachability cannot be removed, make reachability insufficient: require authenticated, mutually verified connections between the two workloads so a pivot needs a credential rather than a route. This is a compensating control, not a chokepoint; it changes what the intruder must steal, and it produces no deny you can point at. ## Writing the residual, which is the part candidates skip Whatever you choose, one sentence has to end up somewhere a risk owner reads it. Something like: *guest-to-guest traffic between workloads co-resident on a provider-operated hypervisor is enforced only by agents inside the guests; an intruder holding local privilege in either guest can disable the near-side agent, and no control we operate is in the path.* Attach who accepted it, what would change the answer (provider feature, dedicated tenancy, workload separation), and when it is reviewed. An architect who ranks the options and states the residual is doing the job; one who says "we use microsegmentation" and stops has not answered the question. ## How to structure the spoken answer Name what you own, name the three positions, rank them against the specific pivot, quote one real cost for each, then state the residual and its owner. Roughly ninety seconds, and it demonstrates the thing the leaf exists to test: choosing a chokepoint is choosing which blind spot you are prepared to live with.