Information Security and AI Use Policy
Service: Doxy (doxy.hu)
Operator: DG-Star Szolgáltató Kft. (company registration number: 13-09-205295, tax number: 27920901-2-13)
Version: 1.0
Effective: 17 November 2026
Approved by: Dános Bálint, founder of Doxy
Next review: no later than 17 November 2027
1. Purpose and scope of the policy
This policy sets out how Doxy protects the documents and data uploaded by its customers, and how it uses artificial intelligence (AI). The policy covers the whole operation of the Doxy service: development, operations, suppliers, AI providers and every person acting on behalf of Doxy.
The policy follows the domains of the Cloud Security Alliance AI Controls Matrix (AICM). Doxy refers to this document in its CSA STAR for AI Level 1 self-assessment.
The policy applies alongside Doxy's General Terms and Conditions (ÁSZF), the Privacy Notice and the Data Processing Agreement (DPA). In the event of a conflict between them, the contract concluded with the customer prevails.
2. Organisation, roles and responsibilities
Doxy is currently a founder-led business. The founder is responsible for all information security, data protection and AI risk management tasks:
- approving, publishing and annually reviewing the policies,
- recording and managing risks,
- managing access to production systems,
- handling and reporting security incidents,
- selecting and reviewing suppliers.
At present, only the founder has administrator access to the production systems. A new employee or subcontractor may be given access only if they have first signed a confidentiality statement, have read and accepted this policy, and they receive only the permissions their task requires (see point 15).
3. Compliance and legislation
Doxy keeps a record of the requirements that apply to it and reviews them at least annually and whenever there is a significant change. At present these are:
- the EU General Data Protection Regulation (GDPR),
- Act CXII of 2011 on the Right of Informational Self-Determination and on Freedom of Information,
- the Hungarian rules on the retention of accounting records and on electronic invoicing (because of the NAV add-on),
- the transparency obligations of the EU Artificial Intelligence Regulation (AI Act),
- the contracts concluded with customers (ÁSZF, DPA).
4. Risk management, including AI risks
Doxy maintains a risk register that contains the description, likelihood, impact and owner of each risk, together with the treatment and its deadline. The register lists AI-specific risks separately, at least:
- an incorrect or fabricated AI answer,
- a document filed to the wrong client or as the wrong type,
- an instruction hidden in a document (prompt injection),
- data leakage between customers (firms),
- an outage of an AI provider or a change in its behaviour,
- an unexpected increase in AI costs.
The register must be reviewed at least annually, and before every significant change (a new AI model, a new provider, a new feature). Before introducing a new AI feature, Doxy carries out an AI impact assessment covering the data protection, security, legal and operational impacts.
5. Classification of assets and data
Doxy keeps a record of its assets: cloud projects, source code repository, provider accounts, data stores, endpoint devices.
Data classification:
| Category | Example | Handling |
|---|---|---|
| Strictly confidential | Documents uploaded by customers, extracted data, AI questions and answers, access secrets, NAV access credentials | Stored encrypted, accessible only to the members of the firm concerned, access logged |
| Confidential | Account data, billing data, logs | Accessible only to authorised persons |
| Internal | Source code, internal documentation | Accessible only to Doxy |
| Public | Website, blog, public policies | May be published freely |
Every customer document is strictly confidential by default, regardless of its content.
6. Access management
Customer access
- Every user has an individual account; shared accounts are not permitted.
- A firm's data can be accessed only by the firm's invited members. The database enforces this on every table with row-level security (RLS), which fails closed in the event of an error.
- Each firm has one owner, who holds every permission; the permissions of the other members are set person by person (for example viewing documents, exporting, deleting documents, viewing the log), and their access can also be limited to particular clients. The team is managed by the owner and by members holding the "invite members and manage their permissions" permission; nobody can grant more permissions than they hold.
- Two-step authentication (MFA) is available to every user, and the owner can make it mandatory for the whole firm. Setting up a NAV connection requires a two-step sign-in.
- The owner can review the firm's users and their permissions in Settings at any time, and everyone can see their own sign-ins.
Doxy's own access
- Every provider account (hosting, database, code repository, AI providers, payments) is protected with MFA.
- The database's service key exists only on the server side, as an environment variable; the browser never receives it.
- The platform administrator role is separate from the customers' roles, and admin operations are recorded in a separate log.
- Doxy looks into a firm's documents only at the customer's request, for customer support purposes. It logs this.
- Doxy reviews its own access annually and revokes any that is no longer needed.
- When an employee or subcontractor leaves, their access must be removed on the same day.
Passwords and authentication data
- We store passwords only in hashed form (Supabase Auth). Password reset takes place via a confirmed e-mail address.
- The number of reset and MFA recovery attempts is limited.
- Doxy stores its own passwords in a password manager, each of them unique.
7. Encryption and key management
- Every data connection runs over TLS 1.2 or a newer protocol.
- The database and file storage are stored encrypted at rest with AES-256 encryption (Supabase / AWS).
- The NAV technical user's password is never stored, only its hash; the hash and the signing key are stored separately encrypted with AES-256-GCM, whose key exists only in the server environment. Neither the user interface nor the database API can read them, and they are decrypted only for the duration of a NAV query.
- Secrets (API keys, webhook secrets) are never placed in the source code; they exist only among the environment variables of the environment that needs them.
- The application's secrets must be rotated at least annually, and immediately if a leak is suspected or a person with access leaves.
- The keys for encryption at rest are managed by the hosting providers; their management is covered by the providers' independent certifications.
- A change affecting encryption may be introduced only through the usual change management process, after assessing its impact on the stored data.
8. Secure development and change management
- Every change is made on a separate branch and reaches the production environment only after a successful build, type check and code analysis (lint), and after the tests that verify the change have run. The full unit, database and end-to-end test suite runs every week, and for every change whose risk calls for it.
- Database changes are committed to the code repository as versioned migrations.
- Live customer data never enters a development or test environment. Tests run on a separate project, with synthetic data.
- Every release can be rolled back to the previous working version.
- An urgent (security) fix also goes through the above process, at most with an accelerated test run. The exception and its reason must be recorded in the description of the change.
- Doxy continuously monitors vulnerability alerts for external libraries and installs the fixes after testing.
9. Logging and monitoring
Doxy logs at least the following events: sign-in, MFA setup and verification, new device, document upload, view, download, export and deletion, AI questions, billing changes, administrator operations.
- A log entry contains the time, the firm, the user, the action, the item concerned and the IP address.
- The content of documents is not written to the logs.
- The application cannot modify or delete the logs. A firm's log can be seen only by its owner and by members holding the "view log" permission.
- Errors, provider outages and failed background processes send an alert to the founder.
- The metadata of every AI call (firm, document, model, token usage, cost, result) is logged.
10. Vulnerability and threat management
- Doxy maintains a threat model that covers AI-specific threats (according to the OWASP Top 10 for LLM Applications), and updates it whenever there is a significant change.
- It prioritises vulnerabilities by CVSS severity. Remediation deadlines: critical 48 hours, high 7 days, medium 30 days, low the next scheduled release.
- It tracks fixes in the code repository. It receives automatic notifications of new vulnerability alerts.
- A software bill of materials (SBOM) of the dependencies is generated automatically with every release.
- Doxy regularly and automatically compares the production database schema with the versioned migrations, and receives an alert if there is a discrepancy.
- The system sends an automatic alert about unusual volumes of downloads and exports.
- An independent penetration test may be commissioned once the number of customers justifies it.
- The system checks the type and size of uploaded files. Files are accessible only through short-lived signed links and are never executed.
11. Handling security incidents
Severity levels
| Level | Example | Response |
|---|---|---|
| Critical | Unauthorised access to or leakage of a customer document, data flowing between firms | Immediately |
| High | Service outage, suspected leakage of an access key | Within 4 hours |
| Medium | Partial feature failure, unsuccessful attack attempt | Within 1 working day |
| Low | Other security observation | By the next release |
Process
- Detection and recording: every incident must be recorded in the incident log (time, description, parties affected, severity).
- Containment: closing off access, rotating keys, shutting down the affected feature.
- Investigation: establishing the cause and the impact, preserving the evidence.
- Notification: in an incident involving personal data, the customers concerned must be notified without undue delay. Where the GDPR requires it, the Hungarian National Authority for Data Protection and Freedom of Information (NAIH) must be notified within 72 hours.
- Closure and lessons learned: corrective measures, updating the policy as needed.
Doxy reviews the incident log at least annually to identify recurring causes. Every quarter it records the security metrics: the number of incidents by severity, the resolution time, the open vulnerabilities and those fixed within the deadline, and the anomaly alerts. It records key rotations, access reviews and exercises in the same security log. Doxy keeps a record of the contacts at the authorities: NAIH, the National Cyber Security Institute (Nemzeti Kibervédelmi Intézet), the police's cybercrime unit. It checks the contacts annually.
12. Business continuity and backup
- The database is backed up automatically every day, and point-in-time recovery is possible.
- A geographically separate (off-site) copy of the stored files is made.
- Transitional state: the daily backup and the off-site copy have not yet been introduced. Doxy will switch them on no later than before the first paying customer's live documents are uploaded. Until then, the data is held on AWS's durable, redundant storage, and customers can export it at any time.
- Doxy tests recovery at least annually and records the result.
- If an AI provider has an outage, processing pauses automatically, the documents wait in a queue, and processing resumes once the service is restored. A fallback model is also available.
- Customers can export all of their documents at any time.
- In case the founder becomes unavailable, the critical access credentials are securely held in escrow by a trusted person.
- In the event of an outage or incident, Doxy informs customers by e-mail and within the application.
- Doxy reviews and exercises this point and the associated recovery plan annually.
13. Suppliers and AI providers
Doxy uses only suppliers that provide a data processing agreement (DPA) and hold an independent security certification or audit report. Its current subcontractors (sub-processors) are listed below; the current detailed list (contracting entity, data handled, legal basis of transfer) is on the Sub-processors page:
| Provider | Task | Data location |
|---|---|---|
| Supabase (AWS) | Database, file storage, authentication | EU, Frankfurt |
| Vercel | Running the application | EU, Frankfurt (fra1) |
| Anthropic | Reading documents, AI answers | United States, on the basis of standard contractual clauses (SCC) |
| OpenAI | Text embeddings (for search) | United States, on the basis of SCC |
| Inngest | Scheduling background jobs (identifiers and processing state only) | United States, on the basis of SCC |
| Sentry | Error monitoring | EU (Germany) |
| Stripe | Payments | EU (Ireland) and United States |
| Resend | Sending and receiving e-mail | United States |
| Billingo | Issuing Hungarian invoices | Hungary |
- Doxy does not use any AI provider that uses customer data for model training.
- Only the document content needed for processing is passed to the AI provider.
- Before a new supplier is engaged, a risk assessment is carried out, and customers must be informed in accordance with the DPA.
- Doxy reviews its suppliers, their contracts and their certifications annually.
Shared responsibility. The physical data centres, the network and encryption at rest are the responsibility of the hosting providers. The application, access management, separation between firms and the use of AI are Doxy's responsibility. The users within the firm, the permissions, switching on MFA and the lawful handling of documents are the customer's responsibility.
14. Data protection and data handling
- The customer (the firm) is the controller of its own data and that of its clients. Doxy is the processor.
- Doxy applies data protection by design and by default: storage in the EU, data-minimising AI calls, no third-party tracking in the application.
- Doxy never deletes a document automatically. Deletion is always a human decision. A file deleted by the firm is also permanently deleted from storage.
- After the subscription ends, the customer can export its documents. The data must be deleted on request within the period set in the DPA.
- Data subject requests (access, rectification, erasure) are fulfilled by the firm acting as controller; Doxy assists with this.
- Before any new processing involving personal data, Doxy carries out a data protection impact assessment (DPIA).
- In the event of a request for data from an authority, Doxy checks its legal basis, discloses only the data that is strictly necessary and, unless prohibited by law, notifies the customer concerned.
- All data extracted from documents and every AI answer refers to the source document and the page.
15. Employees, subcontractors and working environment
This point applies to every person who has access to Doxy's systems, including the founder.
- Before access is granted, a confidentiality statement must be signed and this policy must be accepted.
- Before a new employee is hired, a background check is carried out to the extent permitted by law.
- Every person with access receives security and AI use briefing annually.
- On leaving, company devices must be returned and access must be removed on the same day.
- The screen of a device left unattended must be locked. Confidential data must not be left open or printed out.
- When working remotely, work may be done only from an encrypted device, on a trusted network. Customer documents must not be saved to a local device.
16. Endpoint devices
- A record of company devices is kept.
- Mandatory: full-disk encryption, automatic screen lock, firewall switched on, built-in malware protection, automatic system updates, and the ability to locate and wipe the device remotely.
- Only software from an official store or from the manufacturer may be installed.
- Before disposal, the device must be reset to factory settings. Because of the encryption, this makes the data unrecoverable.
17. Acceptable use of AI and human oversight
- Doxy uses AI exclusively for recognising and classifying documents, extracting data, summarising, and answering questions about the documents. It does not make decisions about natural persons.
- Doxy does not train or fine-tune models, and does not run any AI agent that would carry out actions autonomously or access external systems.
- The content of documents is passed to the model as data, not as instructions. The system validates the model's output against a strict schema.
- Uncertain classifications go to human review ("Rendezetlen" (Unsorted) list). Users can correct any AI decision.
- Every AI answer indicates its source. The ÁSZF states that AI can make mistakes and that answers must be checked.
- Doxy documents the models used and their tasks, and tests them before changing models.
- Customers may use the service only lawfully and for their own documents (ÁSZF).
18. Portability
Customers can export all of their documents at any time in their original format, together with the associated data. Import and export take place exclusively over an encrypted connection.
19. Audit and review
- Doxy carries out and updates the CSA STAR for AI Level 1 self-assessment (AI-CAIQ) annually.
- Doxy keeps a record of the findings of checks and internal reviews. It assigns a corrective measure, an owner and a deadline to each.
- Doxy commissions independent external assessment (penetration test, certification) based on the risks and the number of customers.
20. Exceptions
Deviation from the policy is permitted only in justified cases, with an exception recorded in writing. The exception contains the reason, the risk, the compensating protective measure and the expiry date. The exception is approved by the founder. Exceptions must be re-evaluated at the annual review.
21. Review
The policy must be reviewed at least annually, and whenever there is a significant change (a new provider, a new AI model, a new feature, an incident, a change in legislation). Amendments must be published on this page together with the version number and the date.
Version history
| Version | Date | Change |
|---|---|---|
| 1.0 | 17 November 2026 | First release |