The platform team in this story does not exist. Its details are invented to keep the lesson tidy, and any resemblance to your own team is a coincidence you may want to look into. The team runs a mid-sized product on a public cloud: forty-odd accounts, six squads, one tired platform group. For two years they believed security was handled because their build pipeline scanned every change. Then a Thursday afternoon taught them otherwise, and the next ninety days became a small, practical lesson in CloudSecOps.
The Thursday that changed the plan
A support engineer noticed that a link to an exported report opened without a login. The report sat in a storage area that, it turned out, had been made public eight months earlier during a debugging session and never closed. Nothing sensitive was in it, which was luck, not design. The pipeline had never flagged it because the setting was changed by hand, outside the code. The team’s lead asked three questions in the retro: who owns that storage, why did we not see the change, and how many more like it are out there?
Nobody could answer any of them. That silence was the finding.
Days 1 to 30: see everything
The team started by borrowing a plan from the Ops Atlas page on CloudSecOps, which sketches a 30, 60 and 90 day path. The first month was only about visibility:
- They turned on organisation-wide audit logging and sent it to a separate archive account that no squad could modify.
- They ran a benchmark scan of every account and wrote down which release of the benchmark they used, because the Atlas warns that scanning tools sometimes trail the newest version.
- They listed every encryption key and every long-lived access credential, with the team that created it.
The scan produced hundreds of findings. Nobody panicked, because the aim of month one was not to fix them but to know them. The team discovered, with some embarrassment, that their activity history was only kept for about ninety days by default and that detailed object-level records had never been enabled.
Days 31 to 60: put names on things
Month two was about ownership. The platform group sorted findings by exposure, with public storage and unprotected administrator logins at the top, and gave each critical item to a named squad with a due date. The rule they adopted came straight from the Atlas: every serious finding needs an owner within one working day, and a fix counts only when a re-scan confirms it.
They also created an exception register. A squad that could not fix something in time could ask for an exception, but it had to name a reason, an owner and an end date no later than ninety days out. The first review meeting was a little tense. By the third, squads were closing items just to avoid appearing on the weekly list.
Finally, they tried their first preventive rule: blocking public access on storage. They did not deploy it everywhere at once. The Atlas notes that one major platform offers no evaluate-only mode for these organisation policies, so they applied the rule to a small test group first, watched for a week, then widened it. A broken deployment was caught in the test group, which is exactly why they had one.
Days 61 to 90: enforce and rehearse
In the last month the rules went out account group by account group. The team also ran its first game day. The scenario was blunt: an engineer’s login is stolen and used to create new access keys. Could the logs reveal it? They could, mostly, though one account lacked the detailed records. The exercise also exposed a nasty detail from the Atlas notes: shutting down the stolen session was not enough, because the intruder could assume a role again and temporary credentials last for hours. The runbook was changed to close the re-entry route first.
The team ended the quarter with a one-page scorecard. Coverage was simple division: 38 accounts under the current rule set out of 41 in use, or 92.7 percent, which named the three stragglers outright. They reported the median time to close critical findings next to the oldest open item, so the stragglers stayed visible.
The part nobody enjoyed
It would be tidy to say everything went smoothly. It did not. Two squads pushed back on the owner-within-a-day rule, arguing that they were already stretched. The platform group answered by shrinking the ask: the squad had to claim the finding within a day, not fix it, and the fix deadline scaled with severity. Another squad kept resubmitting the same exception. The group applied the Atlas yardstick for regression, treating a renewed-but-never-closed exception as a failing item, and escalated it to the squad’s manager after the second renewal. Friction like this is normal, and it is a sign the process is biting.
There was also a tooling lesson. One scanner reported a control as passed because a finding had been suppressed during an earlier clean-up. The team caught it only because they compared the pass rate against a manual spot-check of five accounts. After that, they stopped quoting pass rates in meetings and talked about coverage and time-to-close instead.
What they would tell the next team
- Do not start with a blanket restriction. The team considered one and dropped it after reading about teams who triggered outages and then granted wide exceptions.
- Be suspicious of green scores. A control can show as passed when findings are suppressed.
- Keep console edits out of production, or capture them in code the same day.
- Rehearse. Calm in the real incident was earned in the practice one.
A sister team facing the same period later compared notes on running operations during fast-moving incidents, and found useful parallels in a piece on handling incident response with automation and in a practical outline of the skills modern operations professionals need.
Looking back, the biggest change was cultural rather than technical. Squads began to ask the platform group questions before making risky edits, because the audit trail meant surprises would surface anyway. Asking first turned out to be faster than explaining afterwards, and the weekly list got shorter every month.
The Thursday storage link is closed now. More importantly, the team can answer all three of their retro questions, and every one of those answers lives on a page with a date on it.

