Book a demo

Itzal Guard

The prompt box, checked on the laptop

Ten AI tools, four browsers, eight kinds of sensitive data. Guard reads what is about to be sent on the employee's own machine and warns or stops it there, the text itself never reaches us.

Book a demo See pricing

  • 10AI tools covered
  • 8detectors on the device
  • 2operating systems

What you get

What a security team can do differently

Not features, the things a security team can actually do differently once Guard is on the fleet.

  1. 01

    A list of who is sending what, to which AI tool

    Every finding, with the tool, the framework it falls under and the action taken. The sensitive text is masked before the record leaves the laptop.

  2. 02

    A warning that reaches the person in time to matter

    An inline banner in the AI tool's own page, before the send. They decide, and the decision is recorded.

  3. 03

    A block you can turn on per framework, not per person

    HIPAA can block while Internal only warns. You set it once and it applies to the fleet.

  4. 04

    An inventory of the AI tools your staff already use

    Turned on per tenant. It reports the hostname of a tool nobody put on the list, never the content, which is how you find out what you are actually dealing with.

Where it works

The tools people actually paste into

Capture is built per tool. A tool outside the list is still reported by hostname if you turn discovery on, and never by content.

  • AI tools, in the browser

    ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity, Grok, DeepSeek, Le Chat, Meta AI and Poe.

  • Browsers

    Chrome and Edge on MV3, plus Firefox and Safari. Pushed by browser policy through your MDM, never installed by the employee.

  • Desktop apps

    The ChatGPT, Claude and Codex desktop apps on Mac and Windows, read through the operating system rather than by intercepting the app's connection.

Frameworks

Every framework brings its own thresholds and its own action

A finding is not just a kind of data, it is a kind of data under an obligation. Each framework carries its own thresholds and its own action, which is what lets one laptop block for HIPAA and only warn for Internal.

  • HIPAA

    Health records

    Patient names, record numbers and clinical detail, under the identifiers the rule lists.

  • PCI DSS

    Card data

    Card numbers validated by Luhn and BIN, so a test number next to the word example does not fire.

  • GDPR

    EU personal data

    Names, addresses, identifiers and the context that makes them a record rather than a mention.

  • CCPA

    California residents

    The same personal data, under the obligations that apply in California.

  • GLBA

    Customer financials

    Account and routing numbers, balances, and the material non-public information around them.

  • FERPA

    Education records

    Student records and anything that identifies a student inside them.

  • SOX

    Financial reporting

    Unreleased results, forecasts and the working papers behind them.

  • Internal

    Your own rules

    Your patterns and your own kinds of thing, in your words: project codename, claim handler, deal name.

Plans

How it is sold across the editions

Guard is in every edition. What changes is how far it reaches: Guard Essentials covers the AI websites staff use in a browser, and Guard Pro adds the desktop apps, redaction and tool approval.

  • Guard Essentials Included

    The AI websites staff use in a browser, with log and warn.

  • Guard Pro Included

    Desktop AI apps as well, plus redaction and AI tool approval.

  • Itzal Business Included

    As Guard Pro, alongside Shield and Mail on the same agent.

  • Itzal Enterprise Included

    As Itzal Business, plus text in images and Purview labels.

See the full comparison and prices

What it catches

Ordinary days at work

None of these people are doing anything wrong. That is the whole problem.

  1. Support

    A support agent pastes a customer record into ChatGPT to draft a reply.

    The name, address and account number are found before the send. The agent gets the warning; the record reaches the console with the number masked.

  2. Engineering

    An engineer pastes a config file into Claude to debug it, credentials included.

    The API key detector fires on the secret, not on the rest of the file. Under Internal that warns; you can set it to block.

  3. Recruiting

    A recruiter pastes a CV into Gemini to summarise it.

    Person name, email, phone and postal address, all four at once, which is what makes it a record rather than a question.

When it finds something

You choose what happens next

Three settings, chosen per compliance framework rather than per person, so HIPAA can block on the same laptop where Internal only warns. Most companies spend the first couple of months on the first one.

  1. Quietest

    Make a note of it

    Nobody is interrupted. You get a picture of what is happening before you change anything about how people work.

  2. In the middle

    Tell the person

    A banner before the send, saying what was spotted. They decide, and the decision is recorded.

  3. Firmest

    Stop it going

    Held until the data is taken out. Only where you have decided it is worth it, so the rule does not push people somewhere worse.

Set it to watch for a month, then decide. That is how almost everyone starts.

Book a demo How the rollout works

What Guard does not do

The limits, said plainly

Every product on this site has a boundary. Ours are worth knowing before you buy rather than after, and the console reports them the same way.

It does not read everything your staff type.

It reads the box of a covered AI tool, at the moment of a send. Not keystrokes, not the clipboard, not the screen, not pages merely visited.

It does not cover personal devices.

Guard is on the laptops you manage. A phone somebody owns is outside a device-based control, and we would rather say so than imply otherwise.

It does not stop someone retyping it.

A determined person can read a record on one screen and type it into another. Guard is a control on the ordinary route, not a guarantee against intent.

Questions

What people ask before they book

Short answers. The longer versions are in the pages linked from each one.

Does Itzal see what my staff typed?

No. Detection runs on the laptop. What reaches the console is the finding, the kind of data, the count, the tool, the action, with the text masked. There is no field in the record that holds the message.

What happens if detection is slow?

The send goes through. Guard fails open at 250 milliseconds, because a DLP tool that freezes somebody's browser gets uninstalled by the person it was protecting.

Does it cover the desktop apps as well as the browser?

Yes, on Mac and Windows. The desktop apps are read through the operating system's own accessibility layer rather than by intercepting their connection, which those apps pin and which interception breaks.

Can staff turn it off?

No. The extension goes out by browser policy through your MDM, so it is not in the employee's list of extensions to remove. The agent runs as a managed service.

What about an AI tool you have not built capture for?

If you turn discovery on, you get the hostname and nothing else, never the content. It is how you find out which tools your staff actually use before you decide what to do about them.

Does it work if the laptop is offline?

Yes. Detection is local, so warnings and blocks still happen. Events sit in an encrypted buffer on the device and upload when the network returns.

The rest of the family

Everything else is a licence, not another rollout

Most companies start here, then add the others as licences. There is nothing further to install when they do.

See Guard on your own laptops.

Guard is the entry product. Half an hour, your hardware, and a straight answer on what it would cost.

Book a demo See pricing