Every multi-cloud estate has two networks. One lives in code: the route tables, security groups, firewall rules, DNS zones, and interconnect settings declared in Terraform, Pulumi, or a pile of reviewed change tickets. The other lives in the clouds themselves, and it changes a little every week. Someone opens a port during an incident and forgets to close it. A console fix solves an outage and never makes it back into code. A vendor adds a route during a maintenance window. None of it feels like a security event at the time. Months later, in a SOC 2 observation window, it becomes one: the estate cannot show that the network running today is the network that was approved.
Drift is rarely dramatic. That is exactly why it survives. The work is not finding one bad rule. The work is building a routine that notices small differences early, while they are still cheap to explain.
Decide what counts as the baseline, per cloud
A drift check is only as good as the baseline it compares against, and a multi-cloud baseline cannot be one flat list. The same intent, say "only the application tier may reach the database tier on the database port", is enforced by security groups in one cloud, network security groups in another, and firewall policies in a third, each with different defaults and different ways of expressing a deny.
Start from intent, then record the expected enforcement in each cloud in that cloud's own terms. For each cross-cloud segment, write down the expected routes, the expected allowed flows, the expected DNS answers, and the expected logging state. Keep that record in version control next to the code that creates it. When a reviewer asks why a rule exists, the answer should be in the record, not in the memory of whoever was on call.
Avoid baselining everything. An estate wide comparison against every default and cosmetic setting produces thousands of harmless differences and trains everyone to ignore the report. Baseline the controls that matter: ingress and egress paths, route advertisements, encryption settings on interconnects, private endpoint exposure, DNS resolution for internal names, and whether flow logs are actually being delivered.
Read the live estate on a schedule, not on request
Drift detection that runs only when someone remembers is not detection. Poll each cloud's network configuration on a fixed cadence, daily for high change environments and at least weekly otherwise, using read only credentials scoped to that job. Store each snapshot with a timestamp, so a difference can be placed in time rather than argued about later.
Compare in both directions. Code that was applied but never took effect is as much a finding as a manual change nobody declared. A rule that exists in code and not in the cloud may mean a failed apply, a later console edit, or a resource recreated outside the pipeline. Each explanation has a different fix, and the snapshot history is usually what separates them.
Normalise before comparing. Cloud APIs return the same rule in different orders, with different generated identifiers, and with defaults made explicit in one cloud and implied in another. A comparison that flags ordering noise as drift will be muted within a month. Sort, strip generated fields, expand defaults, and compare the normalised form. Reserve human attention for differences that change who can reach what.
Give every difference an owner and an expiry
A drift report with no workflow behind it becomes a weekly guilt email. Route each confirmed difference to one of three outcomes, and record which one was chosen.
First, accept and codify: the change was right, so update the code and the baseline, with the same review a planned change would get. Second, revert: the change was temporary or wrong, so remove it in the cloud and confirm the next snapshot is clean. Third, exception: the difference must stay for a stated reason, for a stated period, with a named owner. Exceptions without an expiry date are just undocumented architecture, and auditors treat them that way.
Incident fixes deserve a special path, because they are the largest honest source of drift. When a responder changes a network control during an outage, the incident ticket should carry a drift task that cannot be closed by the incident itself. The fix stays in place while the incident is open. Within a few days, someone either codifies it or reverts it, and the snapshot history shows both steps.
Watch the quiet controls, not just the rules
The most valuable drift findings are often not in the rules at all. Flow logs that stopped being delivered after a storage change. An interconnect whose encryption setting reverted during a provider migration. A private DNS zone that is no longer associated with one of the networks it serves, so internal names resolve over a public path in one cloud only. A route that is still advertised but no longer preferred, leaving a failover path untested and effectively absent.
Include those states in the baseline: logging delivery, encryption posture, DNS association, route preference, and health check targets. A network can match its rule baseline perfectly and still have lost the evidence trail an auditor needs, or the failover path an incident will need. SOC 2 readiness is about whether controls operate over time, and quiet controls are where operating over time fails first.
Turn snapshots into evidence without extra work
If snapshots are timestamped, normalised, and kept with the decisions made about each difference, the audit conversation changes shape. Instead of reconstructing the past from tickets and memory, a team can show the baseline, the snapshot history, the differences found, and how each one was resolved. That is the shape of change management evidence: not a claim that nothing ever drifted, but a record that drift was looked for, found, and handled.
Keep retention aligned with the observation window. Snapshots deleted after thirty days cannot support a review covering six months. Store them cheaply, in object storage with versioning, and test a restore once. Evidence that cannot be retrieved is a finding waiting for the worst possible moment to announce itself.
A short checklist for cross-cloud drift control
- One intent record per segment, expressed per cloud in that cloud's own terms.
- Baseline only what matters: paths, flows, encryption, DNS, and logging delivery.
- Scheduled, read only snapshots in every cloud, kept with timestamps.
- Normalised comparison, so ordering and defaults never masquerade as drift.
- Every difference codified, reverted, or excepted with an owner and an expiry.
- Incident fixes carry a follow up drift task that outlives the incident.
- Quiet controls checked too: log delivery, DNS association, route preference.
- Snapshot retention matched to the audit observation window, with a tested restore.
No estate stays identical to its code forever, and pretending otherwise is what turns routine change into audit risk. The estates that pass cleanly are the ones that notice their own drift first, explain it while the context is fresh, and keep the receipts as a byproduct of normal operations. That discipline, applied to private connectivity and zero trust controls across clouds, is part of the SOC 2 readiness work described at cloudverg.com, where VergShield and VergMesh are used to keep multi-cloud network controls consistent and observable.
Drift will happen either way. The only choice is who finds it first.
