AI in server maintenance
The AI reads along. Only rules may act.
AI tools increasingly act on their own, often with administrator rights, and can be steered toward someone else’s commands with crafted logs and web pages. With Sentinel, the AI reads and writes; every intervention on your servers is decided by a rule set that you configure and can pause at any time.
- 93%
of 406 companies surveyed in North America had already had an incident caused by AI in their infrastructure in 2026.
- 9 seconds
was all it took for an AI agent, in a publicly known case, to delete a production database along with its backups.
- 0 interventions
can be triggered by the AI. It has no access to your servers, only three read-only checks that Sentinel runs for it.
What the AI does
It explains the plan and writes the report.
- Reads the evidence in serious cases.An unknown account, or a scheduled job that downloads a script: the investigator receives the filtered excerpt and writes what happened and which actions it recommends, with a reasoned confidence level. How it works is described on the security page.
- Explains which vulnerability comes first.One paragraph per server: this vulnerability first, because it is exploited, reachable and has a fix. Order and numbers come from the rule; the AI does the wording.
- Gives the reasons for the maintenance plan.The report says in a few sentences what is due and why.
- Objects when something does not fit.If it considers the order wrong, its objection appears in the audit log and the timeline. That does not stop the run.
- Writes the morning report.Every morning at 08:00, the state of all servers in clear sentences, by email or chat. You set the time.
- Looks at your website on request.After the update, it views a screenshot of your site and notes if something looks broken. Green, yellow or red is still decided by the automated check. This is off by default.
The morning report
How your servers are doing, in four sentences.
You do not have to be an administrator to know whether everything is running. If the AI fails or is switched off, the same status arrives as a list.
All servers are running. One item needs attention.
All 8 servers are running stably. In the last 24 hours, Sentinel successfully completed 5 maintenance runs and installed 23 security updates in the process. An unknown SSH key was discovered on the shop server and should be reviewed. There are no other open items.
What the AI never does
It gets no access and gives no commands.
- It triggers no intervention.The AI has no access to your servers and no tools other than three read-only checks that Sentinel runs for it; updates, blocks and rollbacks follow fixed rules. The investigator may only request more read-only information, at most three times.
- It can be addressed, but not used.Malware now writes sentences aimed at the analyzing AI into log lines and file names. For the investigator, evidence is data between markers, never instructions. An injected sentence can distort the assessment, but can never set or lift a block. A dedicated test with ten deception variants checks exactly that.
- It sees no credentials.It does not get whole logs or package lists; for serious cases it gets a filtered excerpt of at most 80 journal lines relating to the trigger, plus running processes, connections, listening ports, the most recent sign-ins, accounts, sudo rules, newly created file paths and the most recent package operations; all limited in lines and characters. A filter removes recognizable keys and passwords from every text, and IPv4 addresses are truncated.
- It learns nothing from your data.Sentinel trains no model. Every call stands on its own.
- Sentinel does not need it.Without the AI, the same plan is produced and the same maintenance runs; only the explanatory sentences are missing.
Costs and data
The AI has a daily budget, one euro by default.
€1default daily budget
Every call is recorded in its own list with its purpose and estimated cost; you set the daily budget. From live operation: 25 calls cost €0.46 in total.
The AI is Claude by Anthropic, USA. What goes there: server names, titles of maintenance runs and incidents with truncated IPv4 addresses and numbers, for the investigator the excerpt described above, keys only as fingerprints. Screenshots of your website only if you switch that on.
If you do not want to send data to the USA, switch the AI off. Maintenance continues.
During the trial, Sentinel uses LimeByte’s access to the AI. If the control center runs on your side, you can add your own API key if you wish.
The next step
Read your first report.
Six months free of charge in the pilot program, apply by 30 November 2026; the first morning report arrives the day after setup. If you continue: €18.50 instead of €37 per server per month, net, AI included when the control center runs with us, all plans on the pricing page.
Prefer direct contact? +49 221 16834135 · info@limebyte.de