Content never leaves the device
Protection happens at the point of origin, on the employee's own device (the company computer or phone), using proprietary local detection technology, including for text inside images. All that goes to the server is the event record: who, when, which tool, which category of data, which action and, from 01/11/2026, whether the account open in the AI belongs to the company or is personal (for a personal one, the masked e-mail address, such as fm****@gmail.com; the comparison with the company's domains happens on the computer itself); and, if your company turns on AI Use, the daily count of messages per tool and, for each request to an AI, the time, the tool and the work or personal indication (with the subject, for work), computed by a model that runs on the computer itself. The text of prompts, messages, conversations and files is not sent; the only exception is the screen evidence, optional and off by default (see Evidence below). This is an architectural decision, not a configuration option.
On the phone (Android), the checking is done by the Carbon keyboard, installed through the company's mobile device management: the text lives only in the phone's memory, for as long as it takes to decide, and is never stored or sent; only the event record goes out and, in the browser, the website, when it is an AI or messaging site. The phone number is shown masked. In banking apps, the keyboard neither checks nor records anything.
On Linux computers, Carbon only observes and does not block. It notes which programs talk to AI and how much they send, without the content, which is encrypted. It checks on the computer itself what goes to the AI in the sessions of terminal agents such as Claude Code; the AI's answer is never read. It also notes the AI installed on the computer and the credential files within the agents' reach, with the variable names, never the value.
Your data does not go to AI providers
In Carbon, "agent" means the software installed on the computer or phone, not an AI agent talking to an external service. The decision to warn or block is made locally, by the rules your company defines. No data is sent to AI providers or used to train models. Dashboard alerts and signals are also calculated by rules, without artificial intelligence.
Controls
| Area | What we do |
|---|---|
| Location | Data stored in Brazil (São Paulo). Dedicated instance or your own environment on the Enterprise plan |
| Encryption | In transit and at rest, using market standards, with an additional layer for secrets and evidence |
| Segregation | Every record belongs to one organization and all access is filtered by it; dedicated credentials per customer |
| Access | Admin, DPO and auditor roles; two-factor authentication; single sign-on with your company's identity provider |
| Audit | Audit trail of dashboard actions; export to your security monitoring tool |
| Privacy by default | Screen evidence and AI Use off; pseudonymization and record retention chosen by your company in the dashboard; evidence retention configurable (6 months to 10 years) |
| Continuity | Regular database backups; during a cloud service outage, computers already protected keep applying the policy and hold their records to send later |
| Integrity | Records sealed daily with a digital signature and timestamp; any change is detectable |
| Evidence | Optional, encrypted, with the sensitive excerpt covered whenever possible, visible only to your company's DPO and administrator; the Carbon team never accesses it |
| Incidents | Formal procedure, with customer notification within 48 hours |
What Carbon does not recognize yet
- Sensitive data described in free text, without a format of its own: religion, political opinion, racial or ethnic origin, health described in words, sex life, trade union membership. Carbon recognizes data that has a format (documents, cards, credentials, health codes) and the terms your company registers.
- A number shaped like a document but invalid, such as the CPF 123.456.789-00: this is on purpose — the check digit prevents false alarms.
- Password-protected files: the content is not read; the submission is recorded with that mark.
- Text inside images: reading is approximate and may miss data.
- On the phone, an attached or shared file: it does not go through the keyboard and is not checked.
- On the phone, the website when the browser's secure DNS is on: the record shows only the browser. The company's mobile device management can turn that secure DNS off in Chrome.
- On Linux, the website's name when the browser hides it (secure DNS or encrypted name): the row says "unidentified website".
- On Linux, Cursor: its sessions are not read yet.
Sub-processors
| Provider | Use | Location |
|---|---|---|
| Google Cloud | Hosting of the application and database | Brazil (São Paulo) |
| GoDaddy | Alert, invitation and password reset e-mails | United States |
| Cloudflare | Bot protection on the dashboard sign-in page (captcha) | United States (global network) |
Changes to this list are announced to customers 30 days in advance.
Available documents
We send them, free of charge, to any company evaluating Carbon (in Portuguese):
- Privacy Policy (also published here)
- Information Security Policy
- Responsible AI Use Policy
- Access Control Standard
- Retention and Disposal Policy
- Third-Party Management and Business Continuity Policy
- Incident Response Procedure
- Data Processing Agreement (DPA)
- Privacy Principles
- License Terms, SLA and NDA
- Evidence Matrix, DPIA template and employee notice template
Request them at privacidade@carbon-labs.co.
Report a vulnerability
Found a flaw? Write to seguranca@carbon-labs.co. We reply within 2 business days and will not take legal action against anyone who reports in good faith.