1. What the AI agent does not do without a person's confirmation
The AI agent can work in an automatic or an interactive mode, as you configure it. For irreversible actions, in both modes, the agent proposes and a person confirms. Examples: temporarily blocking an account, excluding a type of event from detection, or changing a configuration. The proposed action is shown to a named person, with the buttons “Confirm” and “Cancel”. Who confirmed, and when, is recorded in the audit log, together with everything the agent read and every step it took.
The AI agent never decides whether something is an attack. It explains, using the context, the classification made by conventional detectors that can be tested. At most, a person marks an alert as a false alarm or as wrongly classified.
2. What stays in your network
Logs, detection models, the AI model and the questions you ask the agent stay in your network, in both delivery options. If you work with your own team, nothing reaches certSIGN. In the managed service, certSIGN's specialists work on your platform over a connection you control and can cut at any time from your side, and certSIGN receives only what you choose to send. By default, the platform does not connect to any external service.
One exception depends only on you: the external AI model. If you choose to have the agent use an external AI model through an API, the text of the alerts and their context reach that provider. The option stays off until you switch it on yourself, through a separate contract with that provider, paid by you and under your responsibility. Your choice is written into the installation documentation, where your data protection officer can see it. Detection stays local in every case.
3. What it does not detect or replace
The platform is not an antivirus or endpoint protection, not a firewall, not a vulnerability scanner and not a backup solution. It adds detection on top of the tools you already have, without replacing them. It helps your people; it does not replace them.
The platform works from logs or from a copy of network traffic. A device that produces no logs, on a network segment with no mirror port or tap, cannot be seen. We tell you so before installation, not after. Nothing is installed on your devices, and in the industrial network the platform only listens: it cannot send commands to controllers.
The industrial protocols covered (to begin with, Modbus/TCP and S7comm) are agreed for each site. Before we start, we agree together on the protocols the platform will read on your site.
4. What it cannot guarantee
No detection tool catches absolutely everything, ours included. So that only the alerts that matter reach you, several independent detectors analyse each source and must agree with each other.
When the evidence does not support a conclusion, the platform says openly that it does not have enough information, instead of guessing.
There will be false alarms. When someone on your team marks an alert as a false alarm, the decision is recorded, the detection is adjusted, and the adjustment is recorded too. Tuning is done together and continues for as long as you use the platform.
The reports the platform prepares, including the incident reports NIS2 Article 23 asks for, are drafts. Someone on your team checks each statement against the cited log lines, then approves or changes it. We send nothing to any authority on your behalf.
5. What it does not do for compliance
The cyberTENANT platform does not exempt you from compliance requirements, but it gives you the evidence you need to demonstrate compliance. NIS2 (Article 21) asks for far more than detection. The platform does not replace legal advice, risk analysis or a data protection impact assessment. AI systems used only for cybersecurity are not considered “high-risk” under the EU Artificial Intelligence Act. Even so, we built the platform on the same principles: recorded actions, openly stated limits and mandatory human confirmation for anything that cannot be undone.
6. How to check what we say
- Ask for the audit log. It must show every action the agent took and who confirmed it.
- Ask for the evidence behind any alert. The alert must point to the log lines it started from. If it does not, tell us.
- Ask what changed after a false alarm. You must find both the false-alarm mark and the adjustment made after it, recorded.
- Ask where your data went. The answer must be clear: in your network, except what you chose to send to certSIGN in the managed service and what reaches an external AI model, if you switched one on.
If any answer does not match what this page says, write to us at cybertenant@certsign.ro. If we got it wrong, we correct the page.
Last updated: 5 October 2026.