Detect, intercept and block out-of-policy AI use.

Guard checks what your team sends to AI on each Mac before it leaves, and every block names the rule it hit.

Guard uploads rule ids, match types, counts and, if you allow it, one redacted snippet of at most 80 characters. Never the full prompt. Privacy

For MacsVerified live, 22 October 2026

  • Claude website
  • ChatGPT website
  • Claude for Mac
  • ChatGPT for Mac + Codex

Decided on the laptop, before the request leaves. Not in a dashboard after the fact.

  • 1.2 ms

    to read a prompt, check it against all eight starter content rules and decide, on the Mac. Median for a prompt that is blocked; an allowed one takes 0.8 ms.

    Guard benchmark, Oct 2026
  • 6.5 ms

    from sending a prompt with a tax file number to Guard’s block, end to end through the agent. The AI service never receives it.

    Guard benchmark, Oct 2026
  • 4

    chat apps and websites with their own switch, each checked blocked, allowed and working again in a live run on named client builds.

    Guard live acceptance, 22 Oct 2026

Four controls. One agent on the Mac.

How Guard works

From the rule to the reason.

See AI use

Know which AI services your team uses, and from which apps.

Activity lists every decision: who, which app, which service, which rule. Overview totals services, apps and blocks for the last 7 or 30 days.

What IT sees
  • Chat
  • Coding
  • API
  • Search
  • Writing
  • Image
  • Meetings
  • Agents
  • Notebooks
The Guard console's Activity page with demo data: a sidebar grouped into Protect, Fleet and Catalogue, and a list of decisions (Requested, Observed, Blocked, Asked) with the person and the app, such as Safari, Claude and Google Chrome.

Content rules

Block card numbers, tax file numbers and keys before they are sent.

Rules are deterministic: a checksum or a pattern, matched on the laptop. Block, ask IT, or allow and record. Try a value in the console before the rule ships.

What Guard keeps
  • Card numbers
  • Tax file numbers
  • ABN
  • ACN
  • Medicare
  • Secrets
  • Bulk email
  • Keywords
  • File types
The Guard console's content rules with demo data: tax file numbers, Medicare numbers, payment card numbers and API keys and secrets blocked, ABNs and ACNs record only, and bulk email, bulk phone numbers and spreadsheets set to Ask IT, each on every AI service for everyone.

Switches

Switch off one app or site, not the whole vendor.

claude.ai and chatgpt.com serve both the blocked and the allowed clients, so Guard matches the app’s signed identity and the exact send route, never just the domain. Claude Code and Codex CLI kept working in the live run.

What is covered
  • Claude website
  • ChatGPT website
  • Claude Mac app
  • ChatGPT Mac app + Codex
The Guard console's Rules page with demo data: service rules for DeepSeek (blocked) and ChatGPT, then Apps and websites with four switches: block the Claude website, the Claude macOS app, the ChatGPT website, and the ChatGPT macOS app with ChatGPT and Codex.

Blocks that explain

Every block says why, and what to do next.

The person sees the rule, what was found and a redacted example on the Mac, not a spinner. IT sees the same decision in Activity. An ask goes to the Requests page, and a grant reaches the Mac on its next sync.

What happens on a block
  • Reason
  • Rule
  • Redacted example
  • Request access
  • Monitor mode
The Guard console's detail sheet for a block with demo data: ChatGPT, Blocked, Tax file numbers. Blocked by a content rule, matched once, redacted example 123***782, from Google Chrome to chatgpt.com.

Verified means a live block, allow and recovery run on the named client build. Other browsers, uploads, voice and connectors are not covered; a new client release is re-tested first.

Say yes to AI.

Try Guard on a few of your team’s Macs. The pilot is free for 30 days; to continue, we agree an order form. No automatic billing.

Pilots are by request while we onboard a small group of teams. How a pilot works

For MacsVerified live, 22 October 2026

  • Claude website
  • ChatGPT website
  • Claude for Mac
  • ChatGPT for Mac + Codex

FAQs

Questions IT asks first.

About pilots, installs, what Guard keeps, what happens on a block, how fast it decides, and the free edition.

Talk to us
How does a pilot work?

Pick a small group of Macs and a policy to test. For 30 days, free, we run block, allow and recovery with you, and check content rules on your own workflows. To continue, we agree an order form; there is no automatic billing. Otherwise Guard is removed. Email sam@vibtra.app to start.

What needs an install?

One Mac app. It bundles the agent and installs one system extension: a content filter and a transparent proxy. Enrol with a link from the Devices page; the app creates a certificate for that laptop and the key never leaves it. Nothing is installed in the browser or in the AI app. If the agent is down, the extension fails open and traffic goes straight through, unchecked. Guard runs on macOS 14 or later.

What does Guard keep?

Per event: the rule id, the match type, counts and, if your organisation allows it, one redacted snippet of at most 80 characters. Prompts are inspected on the laptop and never stored by Vibtra. Only catalogued AI hosts are decrypted; everything else is tunnelled untouched.

What does IT see?

A web console with Overview, Activity, Rules, Requests, Devices, People and AI services, plus Settings. Activity shows who, which app, which service and which rule. Every page is admin-only, and Settings holds monitor or enforce and the staff notice.

What happens on a block?

The request stops on the laptop before it reaches the vendor. The person sees Guard’s notice with the rule and the next step; the vendor’s app shows its own generic error. Request access goes to the Requests page, and a grant reaches the Mac on its next sync. New orgs start in monitor mode, which records rule decisions without enforcing those rules. Connections Guard cannot inspect are still blocked unless IT allows them under Uninspectable TLS in Settings. Enforce mode always blocks uninspectable connections.

How fast is Guard, and how were the numbers measured?

On a MacBook Pro (M4 Max) in October 2026, with Guard’s eight starter content rules and a 1.3 KB chat prompt. Reading the prompt and deciding took a median 1.2 ms with a tax file number in it (1.5 ms at p99) and 0.8 ms without. End to end through the agent, with interception and the event record, a block came back in a median 6.5 ms (10 ms at p99), and an allowed request took about the same extra time. These are test results with synthetic prompts on one Mac, not a guarantee for every request or network.

Which apps and sites are covered?

Verified live on 22 October 2026: the Claude website and the ChatGPT website in Google Chrome 156, Claude for macOS 3.1623.1, and ChatGPT for macOS 26.1016 with ChatGPT and Codex as one switch. Content rules apply to every catalogued service the agent can inspect; they are checked separately from the switches, and in a pilot we confirm them on your own workflows. Other browsers, uploads, voice and connectors are not covered; a newer client release is re-tested before it is marked verified. Guard runs on macOS only.

Is there a free edition?

Guard launches with a free Community Edition. It covers the Claude website, Claude for macOS, the ChatGPT website and ChatGPT for macOS, and keeps them current: fixes for vendor changes reach it the same day they reach paying customers. There is no limit on Macs, people or usage. New apps and sites, enterprise features such as SCIM directory sync, hosting by Vibtra and support are paid. Email sam@vibtra.app to discuss them.