A guard on every chart,
five layers behind it
The moment AI reads an email, a web page or a supplier's PDF, someone else's words are in the room. NessisOne assumes some of those words are hostile, and is built so that hidden instructions cannot turn your AI against you.
Three verdicts,
no room for persuasion
Every org chart must include a Security Guard before it can be published. It reports to nobody, so no manager in the chain can overrule it. It only works when outside content is actually read, so it costs nothing the rest of the time.
The content is handed to the seat that needs it, still labelled as outside data.
The content goes to quarantine. You release it or reject it; nothing acts on it until you do.
The content is rejected automatically and recorded in the engagement ledger.
The guard can only answer with one of those three words. It cannot be talked into writing a summary, following a link or "just this once" passing something along. A fixed pattern check runs next to it and can only make a verdict stricter, never softer.
Files you trust can skip screening: you vouch for them once, and every skip is recorded so you can see exactly what bypassed the guard and why.
If one layer misses,
the next one holds
No single check is trusted to be perfect. Hostile content has to get past all five before it can cause harm.
Taint
The moment a run reads anything from outside, every outward action for the rest of that run drops to "asks you first". The taint is saved with the run, so it survives a pause or a restart.
The Security Guard
Screens the outside content itself and returns Clean, Suspicious or Injected, backed by a pattern check that can only tighten the result.
Provenance marking
Outside content is labelled as data, not instructions, and passed between seats by reference. No seat rewrites it into its own words, where a hidden command could hide.
Charter rules
Every seat's standing instructions tell it to report instructions it finds in content, never to follow them.
The confirmation gate
Protected actions need a confirmation the AI cannot give itself. Any confirmation a seat tries to supply is stripped, and the most sensitive actions always go to the owner.
An AI seat never holds
more power than the person behind it
The lowest of three limits
A seat acts on behalf of the person who commissioned the work, and its real authority is the lowest of three things: what its role allows, what that person is allowed to do, and the limit for work running while nobody is watching. A seat can never do something its commissioner couldn't.
Single-use approvals
An approval covers one exact action with its exact arguments, is used once, and a denial always beats an approval.
No break-glass for AI
The administrator override that lets people handle emergencies is withheld from every AI seat, everywhere.
Trust is earned, by hand
A seat with no stated permission asks about everything. More autonomy is only ever granted by a person, based on the seat's track record.
Security is owner-only
Changing security settings is a permission no AI seat can ever be given.
We attack the guard on purpose
The screening bench throws a library of prompt-injection attacks at the models that can staff the Security Guard. In the September 2026 run, four models each caught all 51 attack calls across 17 attack types, and screening cost from about three cents per run.
Protecting what
leaves your company
The guard protects what comes in. These are being built to protect what goes out.
Security Team
In developmentA department the guard can call in for a closer look at text, code or images it is unsure about.
Data Defender
In developmentA seat that checks whether a destination should receive a piece of data before an AI sends it, so a confidential spreadsheet never ends up in the wrong service.
Prompt redaction
In developmentAn opt-in filter that swaps names, numbers and secrets for placeholders before text reaches an AI provider, then puts the real values back in the answer.
Model health
AvailableSentinel quarantines AI routes that stop behaving and fails over in under a second. How Sentinel works
Try to break it
in a live demo
Bring a document you would worry about an AI reading. We will run it through a department and show you what the Security Guard does with it.