Shadow AI is already in your company: your team pastes, drags and attaches company data into ChatGPT, Claude, Gemini and WhatsApp. Carbon shows which AI tools are in use, stops sensitive data before it leaves the computer, and keeps sealed evidence of every decision.
Manus
Lovable
Replit
WhatsApp Web
WhatsApp (app)
Teams (app)
New AI tools appear every week, and teams use them before anyone approves them. The cost of that gap is already measured.
more, on average, per data breach at organizations with high levels of shadow AI
organizations studied had a breach involving shadow AI
of breached organizations had no AI governance policy, or were still writing one
Source: IBM, Cost of a Data Breach Report 2025.
Soon the question won’t be “do you use AI?” but “how do you govern your use of AI?”
One Windows agent watches the four doors data uses today, in the browser and in desktop apps. No proxy, no certificate, nothing to install in your network.
See which AI tools are used and by whom, and decide: approved, tolerated or not authorized.
ManusPasting into the ChatGPT or Claude app gets the same check as the browser, and apps that read your disk show up as AI reads.
A sensitive attachment does not leave through WhatsApp without a decision, on the web or in the Windows app.
Word, Excel, PowerPoint, PDF and zip are read as text, end to end, before they leave.
Carbon turns a blind spot into a decision you can see, explain and prove.
Within minutes the dashboard lists which AI tools your company uses, by person and by area, including the ones nobody approved.
Paste, drag, attach or type: Brazilian IDs (CPF, CNPJ), card numbers with expiry and CVV, credentials and source code are caught before they leave.
Word, Excel, PowerPoint, PDF and zip are read inside, even when the sensitive line is the last one of a 4 MB spreadsheet.
The most common and least watched channel: WhatsApp Web and the Windows app. A sensitive attachment stays on hold until someone decides.
Every decision is recorded and sealed. The DPO exports an evidence package that anyone can verify, without access to your systems.
No keylogger. By default only metadata and a masked excerpt leave the computer. The optional screen capture is seen only by the DPO and the admin.
An international standard for an AI Management System (AIMS), for organizations that develop, provide or use AI.
And Brazil’s LGPD, which asks for a record of every operation with personal data.
| What the standard asks | How Carbon delivers it |
|---|---|
| Clear policies and responsibilities for AI use | A policy per tool and per data type, with admin, DPO and auditor roles. |
| Risk management | Severity per data type, and alerts by e-mail, Teams and Slack when something serious happens. |
| Controls, monitoring and continuous improvement | A live dashboard, trends and the comparison with the previous period. |
| Transparency and traceability | Every use recorded, records sealed every day, and an audit trail of every admin action. |
| A basis to show maturity and trust | A report for the board, CSV for the audit, and an evidence package with a certificate. |
Carbon does not certify your company. It provides the controls and the evidence that an ISO/IEC 42001 or LGPD review asks for.
Source: ISO/IEC 42001:2023.
Without Carbon, a client’s CPF (Brazil’s taxpayer ID) and an AWS key go out through WhatsApp and nobody ever knows. With Carbon in the Windows tray, the message stops before it leaves, the person decides with the reason on screen, and the dashboard records the episode with the data masked. If they decide to send, the DPO sees the screenshot. And if the company marks an AI tool as not authorized, pasting into it is blocked.
Animated demo in four parts. First: on a Windows computer without Carbon, someone types a client’s CPF and an AWS access key into WhatsApp, and both messages go out without anyone knowing. Second: with Carbon in the Windows tray, the same message with the CPF is held at Enter, the Carbon window asks whether the person wants to send it anyway, they choose not to send, and the dashboard records the episode as blocked, with the CPF masked. Third: in ChatGPT Desktop, someone types an AWS key, Carbon asks, the person decides to send, and the dashboard records the episode as justified, with the screenshot for the DPO. Fourth: in the dashboard, the administrator marks ChatGPT (app) as a tool that is not authorized; on the computer, pasting into ChatGPT is blocked, with the Carbon window explaining why, and the dashboard records the episode as a policy event.
Animation with sample data. The Carbon windows show what the agent says when the company policy asks for a justification before personal data or a credential is sent, and when the tool is not authorized. Click the parts above to jump straight to each one.
Each card shows the top 10 and is clickable. Below, the four cards as they appear in the dashboard, with demo data.
By people, with the status you decide: approved, tolerated or not authorized. As a pie chart or a list.
Spreadsheet, PDF, code, presentation. Separates what was sent from what an AI app read from the disk.
AWS key, secret in .env, CPF, card, contact list, source code, and how many were blocked on the device.
People and areas with high-severity episodes. Click and their log opens, already filtered.
Who sent the most, with their area. What left. By day. And below, every episode: time, type, masked excerpt, computer and whether it was blocked. It works for any row of any card.
A Windows installer (Carbon.msi) that you download from the dashboard itself, already carrying your company’s key. Roll it out through GPO, Intune or by hand. Nothing to install in your network: the server is ours.
Within minutes the computers show up and the dashboard starts to fill: pastes into AI tools, uploads, WhatsApp attachments, files read by AI apps, with the excerpt already masked on the device.
Mark a tool as not authorized and the agent starts blocking pastes into it, telling the person why. Approve the official ones. Export the report for the board or the CSV for the audit.
In a tool that is not authorized, Ctrl+V doesn’t happen and the person sees why. It is recorded as a policy event.
Code, keys and personal data hidden in .docx, .xlsx, .pptx and .pdf are detected before upload, even inside a zip.
AWS keys, tokens, secrets in .env, CPF, card numbers, bulk e-mail lists, source code in full or in snippets.
Attachments and pastes in WhatsApp Web; files attached in the desktop apps. The channel that leaks the most is the one nobody watches.
Severity, tool, person and area apply to every card. From the chart to the log in two clicks.
HTML to print as PDF, by section, with the filter in the header. Unlimited, streamed CSV for Excel or a SIEM.
Corporate login (Entra ID), area sync from AD, or a simple CSV. Computers outside the domain too.
E-mail, Microsoft Teams and Slack when a high-severity episode happens. A periodic summary by e-mail.
Only metadata leaves the computer. The excerpt is masked on the device. Configurable pseudonymization and retention. An audit trail of everything the admin does.
A client spreadsheet, payroll, a contract and a zip with code leaving through WhatsApp, with who sent it and their area. Identical rows become one with ×N. Only risks by default, “everything” whenever you want.
Where content went, data moved per tool, use by purpose and by area, the trend, and an inventory with the status of each AI tool.
The difference isn’t a tweak; it’s a new technology. While traditional DLP sifts through network packets, Carbon uses its own AI-guided engine that recognizes the data, the tool and the intent — at the exact point where the leak happens.
| Traditional DLP · reads the network | CarbonAI-guided | |
|---|---|---|
| How it recognizes risk | ✕Fixed rules and signatures on network packets. | ✓Its own AI-guided engine, which understands the data and the intent behind it. |
| The AI era / Shadow AI | ✕Doesn’t know which AI tools the team uses; blind to closed connections. | ✓Maps all the Shadow AI and what leaves through each tool. |
| Privacy | ✕Opens and reads all traffic; installs an intermediate certificate. | ✓Doesn’t read your browsing. Content never leaves the computer — only metadata. |
| WhatsApp and apps | ✕Blind spot in WhatsApp and in desktop apps. | ✓Sees attachments and sends in WhatsApp Web and in the app. |
| Response to risk | ✕Alerts after the data has already left. | ✓Stops it on the spot, in the tool you haven’t authorized. |
| Deployment | ✕Appliances, proxies and certificates inside your network. | ✓One installer, no server in your network, in minutes. |
| Data protection (LGPD) | ✕Sees everyone’s content — invasive monitoring. | ✓Metadata only, with pseudonymization and an audit trail. |
Old DLP reads everything and sees less and less. Carbon, with AI, reads almost nothing and sees exactly what matters.
Free 30-day trial, no credit card. Cancel anytime.
The agent analyzes the content on the computer and sends only metadata to the dashboard: the type of data (key, CPF, code), severity, tool, size and an excerpt that is already masked, like AKIA********TEST. The full text never leaves the computer.
It sees what matters to the company: a file attached or content pasted in WhatsApp Web, and a file attached in the WhatsApp app for Windows, with type, size, severity and the masked excerpt. Conversations are not read or recorded.
No. It’s one Windows installer per computer and the dashboard in the cloud. The agent talks to the server over HTTPS, like a browser. If you prefer a dedicated instance, that’s the Enterprise plan.
Yes. Claude Desktop, Manus and other apps that read files from the disk show up as “AI read”, separate from uploads. WhatsApp, Teams and LinkedIn on Windows are covered for file attachments. In the browser, we cover Chrome and Edge.
The agent starts blocking pastes into that tool and shows the person a notice. Attachments and reads keep being recorded and flagged. The decision applies to the whole organization within minutes.
The dashboard identifies people because that is how you answer “who sent it”. You are in control: pseudonymization, how long events are kept, who has access (admin, DPO, auditor) and an audit trail of everything that was viewed or exported.
You request it below and get a link by e-mail. With it you create your organization and your user, and download the installer. 30 full days, up to the agreed number of computers, no credit card. At the end, the dashboard stays read-only until you decide.
Fill in the form. You’ll get a link by e-mail to create your organization and install the agent.