A preflight check is not protection.
Asking a model to behave is not security. Real enforcement means a protected tool can only be reached through a gateway that re-checks identity, policy and approval at the moment of execution.
NrtVix AI / Company / Security
- The agent proposes
- The Guardian decides
- The gateway executes
- The owner approves what matters
- The audit trail remembers
Three parts. No single point of trust.
The part that talks to customers cannot act. The part that decides holds no credentials. The part that acts only does what was decided.
Agent
Proposes actions. Holds no keys.
Guardian
Decides using identity, policy, budget and the owner’s approval.
Trusted gateway
Executes once, using keys from an encrypted vault.
Audit trail
Every step is recorded and chained.
Assume the agent can be wrong.
Secrets stay out of reach
Tool credentials live in an encrypted vault used only by the gateway. The agent process never receives them.
Decisions are separate
The agent proposes; Guardian decides; the gateway executes. No single part can do all three.
Approval binds to content
An approval is tied to a hash of the exact request. Change the request and the approval is void.
Short-lived, single-use
Permissions expire quickly and work once, so they cannot be saved and replayed.
Tamper-evident record
The audit trail is hash-chained per workspace and can be re-verified at any time.
Fail closed
If a check cannot be completed, the action does not run.
Six checks before anything happens.
- 1
Identity
Is this a registered, active agent in this workspace?
- 2
Validation
Is the tool on the agent’s list, and are the arguments well-formed?
- 3
Policy and budget
Do your limits allow it, and is there budget left?
- 4
Approval
If it needs a person, has the owner approved this exact request?
- 5
Execution
The gateway re-checks everything, then runs the action once.
- 6
Record
A receipt and a chained audit entry are written.
What we keep, and how.
- Workspaces are isolated: one business can never read another’s agents, messages or records
- Agent conversations and decisions are stored so the owner can review them
- Business information you give an agent is reference data only; it can never change what is allowed
- Customer messages are treated as data, never as instructions to the system
- This website is served over HTTPS and collects only what the early-access form asks for
- The hosted service will use encryption in transit and for stored secrets
Security is part of every change.
Automated tests
1,090 tests cover policy decisions, approvals, isolation, the audit chain and failure cases.
Least privilege
Each part gets only the access it needs. The agent gets none to your systems.
No secrets in logs
Keys and tokens are never written to logs, replies or error messages.
Backups before changes
Data is backed up and integrity-checked before every update.
What exists, and what does not yet.
| Item | State |
|---|---|
| Server-side enforcement of every action | Done |
| Tamper-evident audit trail | Done |
| Privacy documentation for US and European customers | In preparation |
| Data-processing agreement | Planned |
| Independent security assessment | Planned before launch |
| Formal certifications | Not started |
Stated plainly. NrtVix AI is in development. It has not yet had an independent security audit and holds no certification. Please do not rely on it to protect real funds until that changes; this page will say when it does.
Found something? Tell us first.
If you believe you have found a security problem in NrtVix AI or on this website, please write to hello@nrtvix.com with the details. We will acknowledge your report, keep you informed, and credit you if you wish.
- Please give us reasonable time to fix a problem before making it public
- Do not access data that is not yours
- Do not disrupt the service for others
Can an agent be tricked into acting?
It can be tricked into asking. It cannot be tricked into acting, because the decision is not made by the model. A request that policy does not allow stays blocked however it is worded.
Where are payment keys kept?
In an encrypted vault that only the trusted gateway can use. The agent, and the model behind it, never see them.
What if the Guardian is unavailable?
Actions fail closed. If a check cannot be completed, nothing runs.
Can audit records be changed?
Changing or deleting an entry breaks the hash chain, and the verification check reports where the chain was broken.
Let agents act. Keep control.
We are opening NrtVix AI to a small group of businesses first. Tell us about yours and we will be in touch.