Security and sovereignty

Security first is not a slogan here. It is the product.

Netarx exists to stop social engineering, deepfake and impersonation attacks. NIK Agent takes that posture and points it at the AI assistant your team is going to use anyway, so the assistant cannot become the next way in.

Sovereignty

Your work stays yours

The frontier labs run your prompts through infrastructure you do not control, on terms you cannot audit. NIK Agent is built the other way around.

Models hosted domestically

We run open-source weights in infrastructure we host in the United States. Your prompts, your files and your agent traces stay in country.

Never trained on your data

Your conversations are not used to train a model. They are not pooled with other customers, and they are not a product.

Yours to govern

Your administrators see your organization's traces and nobody else does. When you leave, you export your conversations and take them with you.
Identity-bound actions

The agent knows who it is working for, and with

NIK Agent is not a chatbot with a login screen bolted on. The Netarx Identity Key is part of how it decides what to do and who to trust.

  • Every action is bound to the person who signed in and to the device they signed in from, so a stolen session is not a free pass.
  • The NIK verdict arrives with the content: green, amber or red, with the reason, on the call, the message or the meeting.
  • The agent treats an unverified instruction differently from a verified one, which is the difference between a wire change and a question.
  • Social engineering, deepfake and impersonation attempts are the threat model the agent is designed around, not an edge case.

The identity layer is the same one Netarx runs on. How NIK works

Green, verified

A known person on an authenticated device. Act as normal.

Amber, check first

A new contact, or signals worth a second look. Confirm out of band before you act.

Red, not them

The models found a fake voice, face, message or image. Stop.

Chain of thought

Reasoning you can read, and review after the fact

An assistant that cannot show its work cannot be trusted with a wire, a contract or a customer. NIK Agent shows its reasoning as it goes and keeps the trace.

Shown before it acts

The reasoning appears with the answer, not in a debug panel, so the person making the decision can see how the agent got there.

Logged for administrators

Traces are retained for your organization on a window you set, so compliance can answer what the agent did and why, months later.

Visible, never hidden

There is no undocumented mode where the agent acts without a trace. If an action matters, an administrator can review it.
Guardrails

Built to be hard to talk into something stupid

The attacks that work against assistants are the same attacks Netarx has been detecting for years: a trusted voice, a false urgency, an instruction from someone who is not who they say they are.

Impersonation in the request

A message that claims to come from your boss is checked against your real relationships and communication patterns before the agent acts on it.

Instructions from content

Instructions hidden inside a document are treated as content to report, not as orders to follow, and the attempt is visible in the trace.

Fake media in the loop

A cloned voice or a face-swapped video that reaches the agent gets the same treatment it gets everywhere else in NIK: amber or red, with the reason.

Actions with consequences

Payment details, new recipients and outbound sends get a second look, and administrators decide how much autonomy the agent is allowed.

Where we are honest

What this page does not claim

A security page is only useful if it is accurate, so here is the boundary.

  • No invented certifications. Certifications, penetration test summaries and subprocessors live on Netarx's trust portal rather than being restated here. trust.netarx.com
  • No perfect detection claim. Detection is a moving target. The product is designed for the miss as well: a decision always carries a non-media fallback, so a person can still confirm out of band in seconds.
  • Sovereignty is about hosting, not magic. If an administrator points the agent at a frontier vendor with their own key, those requests leave our environment and are governed by that vendor's terms. The private model is the default for a reason.

Bring your security team to the first call

We would rather answer hard questions before the rollout than after it. Ask for the security review pack and the deployment plan.