Your people use AI. Your data stays yours.
Tracemark checks prompts and attachments in the browser before they are sent, shows which AI tools are in use across the company, and records who sent what to which tool.
Checked where people actually use AI.
The Tracemark browser extension inspects what is about to be sent, inside the page, with the same detectors as the rest of Tracemark. Only the decision and the evidence you choose to keep reach your server.
Paste and drop
Text pasted or dragged into an AI tool is checked before it lands in the prompt.
Typed prompts
Prompts are checked at the moment they are sent: Enter, the send button or a form submit.
Attachments
Files attached to a prompt are read as far as the browser can, and the policy decides on what it finds.
Files nobody can read
Images, scans and archives can go to your own server for OCR. Only a category comes back, and you decide what happens when nothing can be read.
AI answers
Sensitive data in an answer can be recorded, or blurred until the person chooses to show it.
Any other website
The same inspection can extend to personal cloud, webmail, forums and file sharing, with trusted hosts exempt.
Keep the question. Lose the secret.
Choose per policy what happens when a prompt carries sensitive data.
Block
The prompt is not sent. Point the person to the AI tool your company has approved, or let them continue with a business reason that is recorded.
Mask
The card number, the key or the customer ID is stripped out and the rest of the question goes through.
Warn
The prompt is sent, the person is told what was found, and the detection is recorded.
Monitor
Nothing is shown to the person. The event waits in AI activity for review.
Shadow AI, made visible.
Managed browsers report the AI tools people open, and the agent spots desktop AI apps by window title or process. Every tool lands in one list, and you decide what each one is.
- Sanction the tools your company has approved.
- Block a tool, and every managed browser stops pastes, typed text and uploads on it within minutes.
- Usage by user shows who reaches for unsanctioned tools, with uploads and detections alongside.
- Unknown tools recognised from their traffic are flagged for review before you decide.
- A built-in catalogue of chat, coding, image, writing and meeting tools gives each one a category and a risk estimate.
Who sent what, to which tool.
Each event names the person, the machine, the tool and what the policy did. The text someone entered is kept only when it classifies as restricted or confidential, so ordinary prompts stay private.
- Blocked, masked, warned or logged, plus every time someone continued with a reason.
- Where it came from. See which site pasted text was copied from.
- Redacted evidence keeps the page and the field with every matched value masked.
- Accounts added on AI sites are recorded too.
- Its own permission. Reading what people typed needs a right most roles do not have.
Coding agents, inventoried and governed.
AI coding and automation agents run on developers' machines with access to files, shells and databases. Tracemark lists every installation per user, the MCP servers each one can call, and when it was last used.
- Command line, editor, extension or desktop app, installed or running now.
- MCP servers sorted by what they reach: shell, files, database, browser or a remote service.
- Allowed MCP servers, denied commands and protected files, written into each tool's own enterprise settings.
- Sensitive data in prompts and in files an agent reads, for supported agents.
- Audit first. See what would have been stopped, then enforce.
Every file a browser sent, in one journal.
One row per upload, to AI tools and to every other site the policy watches. An unread file is shown as unread, so it never passes for a clean one.
- AI tools and other sites side by side, ranked by uploads or by bytes.
- Inspection result per file: detected, clean, or not inspected and why.
- Where each file came from, traced to the site it was downloaded from.
- Filters by file type, extension and destination.
- Coverage shown, including machines whose browsers are not reporting.
From a keystroke to a record.
What happens in the page each time someone sends something to an AI tool.
Send
Someone pastes, types, drops or attaches something on an AI site.
Inspect
The extension runs the detectors inside the page, before anything is sent.
Decide
The policy and the data's sensitivity tier pick a response.
Respond
Block, mask, warn or log, with a way forward for the person.
Record
The event lands in AI activity or Uploads, with the evidence you keep.
What is included.
AI Security is a set of modules on one console and one policy engine. Turn on the ones you need.
One central policy for what browsers may send to AI tools. It reaches every enrolled browser at the next sync and can extend to any website.
- Secrets, card numbers, IBANs, national IDs, source code and your own keywords
- Exact match on your own values, stored as hashes
- Block, mask, warn or monitor
- Continue with a recorded business reason
- A decision for files nobody can read
- Observation mode that records without stopping anything
A searchable log of what was typed or pasted into which AI tool, by whom, on which machine, and what the policy did about it.
- User, machine, tool, channel and action per event
- Entered text kept only when restricted or confidential
- Where pasted text was copied from
- Accounts added on AI sites
- Reading it needs its own permission
- CSV export
The AI tools in use across the company, sanctioned against shadow, with the machines and people behind each one.
- A built-in catalogue of chat, coding, image, writing and meeting tools
- Seen in the browser or on the device
- A risk estimate per tool
- Sanction, block or leave for review
- Unsanctioned usage per user
- CSV export
Every AI coding and automation agent on each Windows machine, per user, with the MCP servers it can call and a policy for what it may do.
- Installed or running, and last used
- MCP servers by what they can reach
- Allowed MCP servers, denied commands, protected files
- Sensitive data in prompts and file reads, for supported agents
- Editor extension allow and block lists
- Excluded machines and endpoint groups
Everything in the estate that nobody approved, in one list, built from the sensors that are already running.
- Installed software, and programs that ran without being installed
- Browser extensions, SaaS services and remote access tools
- Unmanaged devices on the network
- Discovered, under review, sanctioned, tolerated or prohibited
- Block from the list, with the reason recorded
- Coverage shown, so an empty list is never read as clean
Every file a browser sent to a site, one row per upload, with what inspection found and where the file came from.
- Top destinations by uploads or bytes
- Filters by file type and destination
- Detected, clean or not inspected, with the reason
- Where each file came from
- Justification shown when a block was overridden
- Coverage of browsers and machines that report
What is checked, and where.
| What | What happens | Where it runs |
|---|---|---|
| Pasted and dropped text | Checked before it lands in the prompt | Chrome, Edge, Firefox |
| Typed prompts | Checked when the prompt is sent | Chrome, Edge, Firefox |
| Attachments | Text and code files read in the browser; images and scans optionally read by your server | Chrome, Edge, Firefox |
| AI answers | Recorded, or blurred until the person chooses to show them | Chrome, Edge, Firefox |
| Desktop AI apps | Discovered by window title or process | Windows |
| AI coding agents | Inventoried, with MCP servers, commands and file access governed by policy | Windows |
Questions.
Does the prompt leave the device to be inspected?
No. Inspection runs in the page, inside the browser. Only the decision reaches your Tracemark server, plus file names and restricted or confidential text if you switch content capture on. The one optional exception: a file the browser cannot read, such as an image, can be sent to your own server for OCR, and only a category comes back.
Will this stop people from using AI at all?
Only if you want it to. Masking removes the sensitive parts and lets the question through. On a block you can point people to the AI tool your company has approved, or let them continue with a business reason that is recorded.
Is everything people type stored?
No. Entered text is kept only when it classifies as restricted or confidential, and only when content capture is on. Ordinary prompts are never stored. Reading the log needs its own permission.
What about desktop AI apps and phones?
Prompt inspection happens in the browser page. Desktop AI apps on Windows are discovered and listed, and AI coding agents are governed through their own enterprise settings. A personal phone is outside what any endpoint product can see.
Can we start without blocking anything?
Yes. Run in monitor mode, or switch on observation mode to record what would have been stopped across every browser. Move to warn, mask or block when you have seen the results.
Works well with
See it on your own AI tools.
We show discovery, prompt protection and the activity log on a test machine of your choice.