Proving your backups for CyFun with your own AI assistant and ECP
Easy Cyber Protection (ECP) helps SMEs and their MSPs reach CyFun and NIS2 readiness. This is a worked example from an SME on CyFun Basic that connected its own AI assistant, like Claude or ChatGPT (here Claude Code), to ECP through the ECP MCP, to prove two backup controls: PR.DS-11.1 (backups are maintained and tested) and RC.RP-01 (the recovery plan is executed). The backup policy said "nightly backups". Reality said otherwise.
1. What counts as backup evidence
Evidence answers one question: can this organisation restore its data when it needs to? Ranked from weakest to strongest:
- → The policy says "nightly backups". It states an intention. It proves nothing about what runs.
- → A "backup OK" screenshot or log line. It proves that a job reported success on one day. A job can report success on an incomplete set, or stop reporting at all. Nobody notices a missing green tick.
- → A restore test with a result. Restore the newest set into a scratch environment, compare it with production, and write down what matched and what did not. This is the only item on the list that shows the data comes back.
A restore test beats a screenshot because it can fail. A screenshot cannot. Keep the failed ones too: a dated failure with a fix is better evidence of a working process than a streak of green ticks nobody checked. Backup basics: the 3-2-1 rule and testing steps.
2. The worked example, including the failed test
The details are anonymised. The sequence is real.
- ECP said what was open. The AI assistant asked ECP what to do next. Two backup controls were open: PR.DS-11.1 and RC.RP-01. The policy said "nightly backups".
- The AI assistant checked the real job instead of writing the evidence text. The nightly database backup had failed silently for 9 days. The export read the data in pages and kept stopping on a size limit, then on "overloaded" errors. A monitoring job had also crashed, on events with an empty organisation id.
- The AI assistant fixed it. Adaptive page size, retry with backoff, and a local copy that rides into the existing off-site backup. It added tests, ran the project gates and landed the change. The next set completed: about 281,000 rows in 119 tables.
- The AI assistant ran a restore test. It restored the newest complete set into a throwaway SQLite file and compared row counts per table with production. Read-only, nothing written back.
The test failed, and that is the useful part
The row counts differed on 11 tables. The cause was not a broken restore. That one set had been taken over 8 days, so rows had been added or deleted between the first table and the last. A backup taken over days is not a consistent snapshot. A "backup succeeded" log would never have shown this. The next step is a set taken in one night, re-tested the same way.
The policy said yes and reality said no. ECP said what to prove, and the AI assistant went to prove it and found the gap. That is the difference between paper compliance and real compliance.
3. What your AI assistant does through the ECP MCP, and what stays human
The MCP (Model Context Protocol) lets an AI assistant you already use work against your ECP organisation. In this example it covered four things:
- → Find what is open. Ask ECP for the next open work and the controls behind it.
- → Check the real system. This part sits outside ECP. The AI assistant reads the backup job, the logs and the storage, with whatever access you give it.
- → Fill fields. Write the document fields: scope, schedule, retention, restore procedure, test result.
- → Attach evidence. Attach the restore test result as text or a link.
What stays with a person
- The review date and the approval. The AI assistant cannot sign off in your name. This is by design: a person who is accountable reads the evidence and approves it in the app.
- Judging whether a failed test is acceptable, and deciding the fix.
- Coverage that has no tool yet. At the time of the example there was no MCP tool for the critical data entities, so that part was entered by hand in the app. A tool for it is being built.
Give an AI assistant read-only access to production where you can, and check what it wrote before you approve. This article is not legal advice, and a passed restore test does not by itself make a control compliant. It is one piece of evidence for the control owner to weigh.
Checklist: run this per client
Once per client, then at the interval the client's backup policy names:
- Open the two controls (PR.DS-11.1, RC.RP-01) in ECP and read what the policy promises: what, how often, how long kept.
- Check the real job: last successful run, last 30 days of results, and who gets an alert when it fails.
- Confirm one copy is off-site and that a failed job reaches a person.
- Restore the newest complete set into a scratch environment. Never restore over production.
- Compare against production: row counts or file counts, plus a spot check of a few records. Write down the differences.
- Check the set is one consistent snapshot, taken in one run. A set spread over days will not match.
- Record the result in ECP, failed or passed, with the date, the restore time and the fix if any. Attach it as evidence.
- Hand the review date and approval to the accountable person. Repeat the test on the schedule.
FAQ
Why is a restore test better than a "backup OK" screenshot?
A screenshot shows that a job reported success. A restore test shows that the data comes back and matches production. It can fail, so a pass means something.
Can my AI assistant approve the control for me?
No. Through the ECP MCP the AI assistant can fill fields and attach evidence. The review date and the approval stay with a person in the app.
Does a failed restore test count against us?
A dated failed test with a fix and a re-test shows a process that works. Whether a control is met is for the control owner and, where relevant, the auditor to judge. This article is not legal advice.