1. Who we are
Carbon is a data leak prevention solution for generative AI and messaging tools, developed and operated by CodeSpark Assessoria Ltda, of São Paulo, SP, Brazil ("Carbon", "we").
Data Protection Officer (DPO): Fernando Oliveira, at privacidade@carbon-labs.co.
2. The principle behind the product
We protect without seeing the content. Carbon acts at the point of origin, on the employee's own device (the company computer or phone), using proprietary local detection technology. All that reaches our servers is the record that an event happened: who, when, in which tool, which category of data and which action was applied.
Content analysis (prompts, messages, files, including text inside images, and what terminal AI agents send to the AI on Linux) happens entirely on the employee's device and is discarded right after: Carbon does not store or transmit that content to our servers. The only exception is optional and under the customer's control: the screen evidence, described in section 4. No customer data is sent to artificial intelligence providers or used to train models, ours or anyone else's.
3. Our role
| Situation | Carbon's role | The other party's role |
|---|---|---|
| Data about the customer's employees generated by using the product | Processor (art. 5, VII, LGPD) | The customer is the controller and defines purposes and rules |
| Business contact data, dashboard administrators, trial requests, billing | Controller | The data subject is the contact or administrator |
When acting as processor, we handle data only according to the customer's documented instructions, under the Data Processing Agreement (CBN-CTR-08, available on request).
4. Data processed as processor
| Category | Description | Notes |
|---|---|---|
| Employee identification | e-mail or network user name, area or department | The area may come from the optional integration with the customer's corporate directory |
| Device identification | computer name, agent version, last contact | Needed for coverage and support |
| Event record | date and time, category of data detected (for example, personal ID, credential), number of occurrences, severity, destination tool or website, browser used, size and type of the text or file, file name, action applied | Does not include the content. File uploads are recorded even when no sensitive data is found |
| Data origin | the program the text came from and, when possible, the short name of the file, the website address or the source folder | Never the content; the folder is recorded without the computer user's name |
| Usage count | per day: which AI tools and WhatsApp the employee used, how many messages were sent and their total size, in characters | Off by default. Only exists if the customer enables it. Never the content, no time of each message and no screen time |
| AI use classification | for each request sent to an AI tool (WhatsApp is not included): date and time, the tool, whether the request looks like work, personal or undetermined, the confidence of that indication and, for work requests, the subject in one of 12 categories (for example, legal, finance, programming) | Off by default. Only exists if the customer turns on AI Use and states the company's industry. The indication is made by an artificial intelligence model that runs on the computer itself: the text of the request does not leave the computer and is not stored, and the model is not trained on customer data. It is a signal, with no automated decision about the employee; the customer's administrator and data protection officer see the figures and, per person, the list of requests with the time, the tool and the indication |
| Account in the AI tool | from 01/11/2026, on each submission to an AI tool (WhatsApp is not included): whether the account open in the tool belongs to the company, is personal or could not be identified and, for a personal account, the masked e-mail address (for example, fm****@gmail.com) |
Read on the tool's screen and compared, on the computer itself, with the company's e-mail domains; the full e-mail address never leaves the computer. Sent with the event record and, if the customer enabled AI Use, with the request's classification. Observation only: it does not block, raise alerts or lead to any automated decision about the employee. With pseudonymization on, the e-mail address is not shown |
| AI agents on the computer | the AI agents configured on the computer: the name of each connection and service, the folders and projects they can access and the name of the credentials they use | Never the value of passwords, keys or tokens; folders are recorded without the computer user's name |
| Masked reference | unreadable fragment that allows auditing without revealing the data | Never reconstructs the original data |
| Decision to proceed | when the customer's policy asks for a justification and the employee chooses to proceed, the record shows that they proceeded (who, when, type of data and destination) | No text is typed or sent. Only exists if the customer enables this rule |
| Signals for review | indications calculated from the records above, such as a possible attempt to get around the protection or a volume of records outside the person's own usual pattern | Calculated by rules, without artificial intelligence; shown to the customer's DPO and administrator roles; meant for human review and never an automated decision about the employee |
| Screen evidence (optional) | image of the screen at the moment the employee confirms a justified submission | Off by default. If the customer enables it: the employee is notified beforehand; Carbon tries to cover the sensitive excerpt in the image, which is not always possible; the image is encrypted, visible only to the customer's DPO and administrator roles (never to the Carbon team) and kept for the period the customer defines, at least 6 months |
On the phone (Android)
If the customer installs the Carbon app on phones owned and managed by the company, Carbon works through the phone's keyboard:
| Category | Description | Notes |
|---|---|---|
| What the keyboard checks | what is typed or pasted in any app on the phone, so that a piece of sensitive data cannot be completed: the key or the paste that would complete it does not go through, and a clipboard holding sensitive data is emptied | The text lives only in the phone's memory, for as long as it takes to decide: it is never stored or sent. The keyboard does not learn words or keep the emojis used. In banking apps, it neither checks nor records anything |
| Event record on the phone | date and time, the app (by name), whether the data was typed or pasted, the category of data, the severity and the masked reference; in the browser, the website, only when it is an AI or messaging site | Never the content |
| Phone identification | the manufacturer, the model, a short device code, the app version, whether the Carbon keyboard and the website reading are on and, when the SIM card provides it, the masked phone number (for example, 11 9***39375), which identifies the person in the dashboard |
The full number is never stored and never leaves the phone |
| Website reading in the browser | to know which website is open in the browser, the app uses a VPN that stays inside the phone itself and only sees address lookups (DNS); browsing does not go through it | Only AI and messaging websites are kept, in the phone's memory and for 10 minutes; other sites are never noted |
On Linux computers
If the customer installs the Carbon agent on Linux computers, it only observes: it does not block or show any window.
| Category | Description | Notes |
|---|---|---|
| Computer identification | the computer name, the Linux distribution, the agent version and the people logged in | Needed for coverage and support |
| Programs that talk to AI | per day: the person, the program name, whether it was started in the terminal or is the browser, the AI website and tool, the number of connections and how many bytes went out and came in | Read only at the start of each connection, never the content, which is encrypted; for the program, only its name, never the command line. Websites that are not AI are never noted; when the browser hides the website's name, the row says "unidentified website" |
| Terminal AI agents | the sessions of agents such as Claude Code and Codex: the project (without the person's folder name), the start, the end and the number of messages; and, when what was sent to the AI contained sensitive data, the event record with the masked reference | What the person typed and what the agent read or ran and sent to the AI are checked on the computer itself, from the history the agent keeps there; the AI's answer is never read, and the text never leaves the computer |
| AI installed on the computer | the AI tools that run on the computer itself (for example, Ollama): the name, the version, the downloaded models and the size of each, whether they are running and whether other computers on the network can use them | Never what was asked of them |
| Secrets within the agents' reach | credential files (such as .env, SSH keys and cloud credentials) in the person's folder and in the agents' projects: the path (without the person's folder name), the type, the names of the variables and whether an AI agent read the file |
Never the value |
The customer can enable pseudonymization in the dashboard, which replaces personal identifiers with non-reversible codes in screens and reports.
5. Data processed as controller
- Dashboard administrators and users: name, e-mail, access role, protected credentials and access logs (date and time, IP address and the result of each sign-in attempt).
- Business contacts and trial requests: name, e-mail, company, phone and message voluntarily submitted.
- Billing data: company name, tax ID, billing contact.
- Website visitors: this site uses no cookies and no audience analytics. The language you choose is stored only in your browser. Fonts are loaded from Google Fonts, which receives the visitor's IP address in order to serve them.
6. Purposes and legal bases
| Purpose | Legal basis (art. 7, LGPD) |
|---|---|
| Providing the contracted service to the customer (as processor) | Defined by the customer as controller, typically legitimate interest or compliance with a security obligation |
| Authenticating administrators and protecting accounts | Performance of contract; legitimate interest (security) |
| Logging access to the application | Compliance with a legal obligation (art. 15 of Brazil's Internet Civil Framework) |
| Answering trial requests and contacts | Pre-contractual steps at the data subject's request |
| Billing and tax obligations | Compliance with a legal obligation |
We do not sell, rent or share personal data for advertising. We do not use customer data to train or improve AI models, ours or anyone else's.
7. Where data is stored
- Storage: an internationally certified cloud provider, in a data center located in São Paulo, Brazil.
- Alternative: dedicated instance or installation in the customer's own environment, under contract.
- International transfers: storage is in Brazil. Ancillary services listed in Annex I of CBN-CTR-08 (such as sending transactional e-mails) may process data outside the country, with the safeguards of art. 33 of the LGPD and ANPD Resolution No. 19/2024.
8. Who we share data with
Only with sub-processors required to provide the service (the public list is on our Security page; Annex I of the Data Processing Agreement, CBN-CTR-08, formalizes it in the contract), and with authorities when there is a legal obligation or court order. Integrations enabled by the customer (corporate directory, security monitoring systems, alert channels) send data to the customer's own systems, under the customer's responsibility.
9. Retention
Event records, the usage count, the AI use classification and, on Linux, the programs that talk to AI and the terminal agent sessions for the period the customer chooses in the dashboard (30 days to 5 years) or, if none is chosen, for the contract term; the installed AI and the secrets within reach show only the current state of each Linux computer; screen evidence for the period defined by the customer, at least 6 months; regular backups; at the end of the contract, data is available for export for 30 days and deleted afterwards.
10. Security
We apply technical and organizational measures that include encryption in transit and at rest, segregation by customer, role-based access control, two-factor authentication and an audit trail.
11. Integrity of records
Event records and evidence are sealed daily with a digital signature and a timestamp issued by an independent certification authority. Any later change is detectable, which gives the records evidentiary value. Customers can verify integrity in the dashboard. The sealing process does not send personal data to third parties.
12. Data subject rights
Data subjects may request confirmation of processing, access, correction, anonymization, blocking or deletion, portability, information about sharing and review of decisions (art. 18, LGPD).
- Data processed as processor (an employee of a customer): we forward the request to the customer, as controller, within 2 business days and support the response.
- Data processed as controller: we respond within 15 days, at privacidade@carbon-labs.co.
13. Incidents
As processor, we notify the customer within 48 hours of confirming a security incident involving personal data, so that the customer can meet its obligations under ANPD Resolution No. 15/2024.
14. Updates
This policy is reviewed at least once a year or whenever there is a relevant change in the product, sub-processors or law. Material changes are communicated to customers 30 days in advance.