Security and compliance
Built for sensitive legal documents
CODEXIS AI processes sensitive legal documents for law firms, corporations and public administration. Security is not a feature — it is the foundation of the entire platform. This page summarizes how we protect your data, which standards we follow and — with the arrival of CODEXIS AI Agent — which deployment models you can choose from according to your risk management policy.
Guarantees
What we guarantee regardless of deployment model
These guarantees apply to every CODEXIS AI customer and to every deployment variant.
Your data does not train the models
Neither prompts nor analyzed documents are used for training — not at ATLAS GROUP, nor at OpenAI. Contractually anchored in the DPA under Art. 28 GDPR.
No access for other customers
Logical per-tenant isolation in SaaS. In an on-prem deployment you are the only tenant.
Ownership stays with you
Inputs, outputs and analyzed documents belong to the customer. Without exception.
Optional E2E encryption (in development)
We are preparing a mode in which the encryption key is known only to the user. After a defined period of inactivity, user data is encrypted and without the key no one can access it, not even ATLAS GROUP.
Zero data retention with OpenAI
We have contractually agreed Zero Data Retention with OpenAI — neither API requests nor responses are stored (zero retention).
EU data residency with OpenAI (optional)
You can optionally enable processing of AI queries within the European Union for OpenAI models. Requests and responses then stay in the EU and do not travel to the USA. You toggle it yourself in the settings (“Data processing”).
Encryption at rest and in transit
AES-256 for data at rest, TLS 1.2+ for all communication.
GDPR and DPA
A template Data Processing Agreement under Art. 28 GDPR is available whenever you need it.
Deployment models
Three models to match your risk management policy
CODEXIS AI Agent can be operated in three basic models. The same core guarantees apply to each; they differ in where the infrastructure sits and who is responsible for what.
Cloud (SaaS)
For most firms and companies
Operated in the EEA (Ostrava + AWS Frankfurt). ATLAS GROUP manages everything — from patches to DR. A fast start with no infrastructure of your own.
Private cloud
For regulated entities and data locality
A dedicated instance in a location chosen by the customer, still managed by ATLAS GROUP. Suitable for regulated entities and clients who need to define their data locality.
On-premise
For regulated entities and data locality
The application layer and primary data can be deployed in your data center or in your cloud, anywhere you choose. Suitable for regulated entities and clients who need to define their data locality. We need Kubernetes and a few other standard components; the detailed deployment description is provided under NDA. AI inference (LLM) runs through the OpenAI API under the same terms as in SaaS.
Model comparison
| Aspect | Cloud (SaaS) | Private cloud | On-premise |
|---|---|---|---|
| Infrastructure operator | ATLAS GROUP | ATLAS GROUP (dedicated instance) | Customer or their cloud provider |
| Application data location | EEA — Ostrava (CZ) + AWS Frankfurt (DE) | EEA — as chosen | Customer data center or a cloud chosen by the customer, anywhere they wish |
| Tenant isolation | Logical per tenant / license ID | Dedicated instance | Single tenant |
| LLM (AI) provider | OpenAI API managed by ATLAS GROUP | OpenAI API managed by ATLAS GROUP | OpenAI API managed by ATLAS GROUP |
| EU data residency (OpenAI) | Optional (opt-in) — AI query processing in the EU | Optional (opt-in) — AI query processing in the EU | Optional (opt-in) — AI query processing in the EU |
| Zero Data Retention (OpenAI) | Agreed — zero retention (DPA) | Agreed — zero retention (DPA) | Agreed — zero retention (DPA) |
| OpenAI BYOK | Possible based on an individual commercial offer | Possible based on an individual commercial offer | Possible based on an individual commercial offer |
| Technical prerequisites | No infrastructure of your own | As agreed for the dedicated instance | Kubernetes + other standard components; detailed deployment under NDA |
| Updates | Managed by ATLAS GROUP | Managed by ATLAS GROUP | ATLAS GROUP delivers, customer applies |
| Role under GDPR | Processor (Art. 28) | Processor (Art. 28) | Processor for the AI path |
| Suitable for | Most firms and companies | Regulated entities and clients requiring a defined data locality | Regulated entities and clients requiring a defined data locality in their own DC or cloud |
Even in an on-prem deployment, AI inference (LLM) is performed through the OpenAI API. The application layer and primary data storage run in an environment you choose: in your data center or in your cloud, anywhere according to your architecture and security rules. From a technical standpoint we need Kubernetes and a few common infrastructure components; the specific deployment description is shared under NDA. Individual AI queries (prompt + relevant context) are still sent to OpenAI under the same contractual guarantees as in SaaS (DPA under Art. 28 GDPR, agreed Zero Data Retention — neither requests nor responses are stored, no training). You can optionally enable EU data residency (the “Data processing” setting), in which case AI queries are processed in the EU and do not leave it. In all deployment models, the OpenAI Bring Your Own Key (BYOK) mode can also be arranged based on an individual commercial offer. On-prem in CODEXIS AI Agent therefore does not mean “air-gapped”; it means residency of the application and stored data at the customer.
Shared responsibility
Who is responsible for what
The division of responsibilities between ATLAS GROUP and the customer according to the chosen deployment model.
| Area | Cloud SaaS | On-premise |
|---|---|---|
| Physical security of the DC | ATLAS GROUP / AWS | Customer |
| Encryption at rest / in transit | ATLAS GROUP (in-built) | ATLAS GROUP delivers, customer operates |
| Infrastructure patch management | ATLAS GROUP | Customer |
| Application patch management | ATLAS GROUP | ATLAS GROUP delivers, customer applies |
| Identity management and SSO | Customer (via IdP) | Customer |
| Data classification and retention | Customer | Customer |
| Backups and DR | ATLAS GROUP | Customer |
| Security monitoring | ATLAS GROUP | Customer (with ATLAS GROUP support) |
| Incident response for the software | ATLAS GROUP | ATLAS GROUP (for SW defects) + customer (operations) |
Certifications
International standards and regulatory compliance
CODEXIS AI is developed and operated by ATLAS consulting spol. s r.o., a company within the ATLAS GROUP — externally we operate under the ATLAS GROUP brand, while contractual documents (DPA, NDA) are concluded by ATLAS consulting spol. s r.o. This company operates a certified information security management system (ISMS). Certificates and audit reports are provided under NDA on request.
| Certification | What it covers | Applies to |
|---|---|---|
| ISO/IEC 27001 | ISMS — information security management | CODEXIS AI development and all deployment models |
| ISO/IEC 27017 | Security controls for cloud services | Cloud SaaS and private cloud operated by ATLAS GROUP |
| ISO/IEC 27018 | Protection of personal data (PII) in the cloud | Cloud SaaS and private cloud operated by ATLAS GROUP |
| GDPR | EU Regulation 2016/679 | All models (with an adjusted division of roles for on-prem) |
| DORA | Digital Operational Resilience Act | Relevant for clients in the financial sector |
Nine pillars
Nine pillars of security
The controls apply across deployment models — they are a property of the software and processes, not of the hosting. Structured according to ISO 27001 Annex A.
- 01
Access and identity management
- SSO via SAML 2.0 and OAuth 2.0
- Integration with Microsoft Entra ID / Azure AD
- MFA enforced at the customer’s IdP level
- Least privilege and separation of duties
- Audit log of all access and role changes
- Regular entitlement review
- 02
Encryption
- TLS 1.2+ for all data transfer
- AES-256 for data at rest
- Encrypted backups
- Centralized cryptographic key management
- Password hashing (bcrypt / PBKDF2)
- HSTS and secure cookies
- 03
Secure development (SDLC)
- OWASP Top 10 in code review
- SAST analysis of source code
- Software Bill of Materials (SBOM)
- Separate dev / test / prod environments
- Version control (git) and code signing
- No hard-coded secrets
- 04
Vulnerabilities and patching
- Weekly vulnerability scan (internal and external)
- Annual penetration test and after every significant change
- Critical / High patches based on CVSS and exposure
- Risk score: CVSS + KEV + exposure + criticality
- Centralized remediation tracking
- In on-prem: ATLAS GROUP delivers patches, customer applies them per SLA
- 05
Logging and monitoring
- Logging of access, authentication and AI operations
- In SaaS: a centralized log host managed by ATLAS GROUP
- In on-prem: logs stay with the customer, export to SIEM
- Tamper-proof logs with integrity
- Retention of at least 6 months (personal data 3 years)
- NTP synchronization to UTC
- 06
Incident management
- A defined process for detection, response and notification
- The ATLAS GROUP security board as the escalation body
- Contractual notification of the customer per DPA
- Post-mortem and root cause analysis
- Compliance with GDPR Art. 33 (72h notification in SaaS)
- 07
Continuity and Disaster Recovery
- In SaaS: geographically separated DR (Ostrava + Frankfurt)
- Regularly tested, ATLAS GROUP manages the backups
- In on-prem: ATLAS GROUP provides a recommended BC / DR architecture
- Business Impact Analysis for every critical asset
- BC plans with defined roles
- SLA with defined availability
- 08
Data governance
- Tenant isolation per license ID (in SaaS)
- Encrypted backups
- Secure deletion after contract termination
- Export of user data on request
- Data subject rights under GDPR fully supported
- Granular access rights at the document, module and agenda level
- 09
AI governance
- No training on your data (contractually)
- Granular deactivation of AI per agenda
- AI respects the user’s application permissions
- Logging of all AI operations
- DPA with OpenAI under Art. 28 GDPR — applies in all deployment models
- Ownership of inputs and outputs stays with the client
AI and your data
Four principles for how AI handles your data
These principles apply across all deployment models.
- 01
Your data does not train any model
Neither prompts nor the content of analyzed documents are used to train or improve AI models — not at ATLAS GROUP, nor at OpenAI. Contractually anchored in the DPA under Art. 28 GDPR and applicable across all deployment models.
- 02
Zero data retention with OpenAI
We have contractually agreed Zero Data Retention (ZDR) with OpenAI — neither API requests nor responses are stored. (Without ZDR, OpenAI would by default temporarily retain data for up to 30 days solely for abuse detection; thanks to ZDR this does not happen.)
- 03
Optional EU data residency with OpenAI
By default, AI queries are processed by OpenAI in the USA (transfer under SCC). For OpenAI models you can optionally enable EU data residency in the settings (“Data processing”) — requests and responses then run in the European region and do not leave the EU. You can turn the option on or off at any time.
- 04
You decide what the AI sees
AI features are standalone modules. They can be completely deactivated, enabled only for defined agendas, or have access to specific sources controlled granularly. AI operations always respect the user’s access rights.
- 05
Ownership stays with you
Inputs are the property of the customer; outputs can be freely used for internal purposes. The AI models remain the property of the provider (OpenAI). The CODEXIS AI provider does not sell user data to third parties.
Subprocessors
A transparent list of subprocessors
The subprocessors involved in running CODEXIS AI. The list changes according to the deployment model — in the on-prem variant, AWS drops out of the main data path.
In Cloud (SaaS) and private cloud
| Subprocessor | Purpose | Location | Contractual guarantees |
|---|---|---|---|
| ATLAS consulting spol. s r.o. | Main operator · application layer · database | Ostrava, CZ · EEA | ISO 27001 / 27017 / 27018 · DPA · NDA |
| Amazon Web Services (AWS) | Cloud / object storage · backups | Frankfurt, DE · EEA (eu-central-1) | DPA · SCC · ISO 27001 · encryption at-rest |
| OpenAI, L.L.C. | AI inference (LLM) — request / response only | USA · transfer under SCC — optionally EU data residency (processing in the EU) | DPA Art. 28 GDPR · agreed Zero Data Retention (zero retention) · no training · SOC 2 Type II |
In on-premise deployment (CODEXIS AI Agent)
In the on-prem variant, AWS drops out of the main data path — the application layer and primary storage run in your infrastructure. OpenAI, however, remains in the path, because AI inference is performed the same way as in SaaS — here too you can optionally enable EU data residency (processing of AI queries in the EU).
| Subprocessor | Purpose in on-prem | Location | Contractual guarantees |
|---|---|---|---|
| ATLAS consulting spol. s r.o. | Software supplier · update channel · remote support · mediation of the OpenAI integration | Ostrava, CZ · EEA | ISO 27001 · narrowed DPA · NDA |
| OpenAI, L.L.C. | AI inference (LLM) — request / response only | USA · transfer under SCC — optionally EU data residency (processing in the EU) | DPA Art. 28 GDPR · agreed Zero Data Retention (zero retention) · no training · SOC 2 Type II |
Frequently asked questions
Questions from security auditors and CIOs
What we are asked during vendor security reviews — from data protection to incident notification and uninstallation.
Do you use our prompts or documents to train AI models?
No. Neither data entered by users (prompts) nor the content of analyzed documents is used to train or improve AI models — not at ATLAS GROUP, nor at the LLM providers. In the SaaS variant this is contractually anchored in the DPA with OpenAI under Art. 28 GDPR. In the on-prem variant the same principle applies.
Where does our data physically reside?
It depends on the chosen deployment model, and AI queries pass through OpenAI: Cloud SaaS — application data in the EEA (Ostrava + AWS Frankfurt), AI query via the OpenAI API (by default the USA, transfer under SCC). Private cloud — application data as agreed, typically in the EEA. On-premise — application data in your infrastructure, AI query via OpenAI. In all models you can optionally enable EU data residency (the “Data processing” setting), in which case AI queries are processed in the EU and do not leave it.
How does the on-prem deployment (CODEXIS AI Agent) differ from the cloud?
On-prem addresses the residency of the application layer and stored data in your infrastructure — the database, object storage, documents and backups run on your servers. Operations, availability, capacity, physical security and infrastructure patch management are provided by you. What on-prem does not address: AI inference still goes through the OpenAI API — by default in the USA (transfer under SCC), with the option to enable EU data residency (processing in the EU). If you need an absolutely air-gapped mode without any outbound traffic, CODEXIS AI Agent does not offer that today.
How is isolation between customers ensured (multi-tenancy)?
In SaaS: logical segregation by tenant / license ID. The application logic verifies permissions on every query — no user of one tenant can access another tenant’s data. Both PostgreSQL and the object storage are structured tenant-aware. In private cloud / on-prem you are the only tenant, and the isolation is inherent to the nature of the deployment.
Do you support SSO and MFA?
Yes, in all models. CODEXIS AI supports SAML 2.0 for federated authentication, OAuth 2.0 / OpenID Connect, Microsoft Entra ID (formerly Azure AD), and MFA enforced at the level of your identity provider. User passwords never leave the customer’s identity provider.
How often do you perform penetration tests and vulnerability scans?
Vulnerability scan: weekly, internal and external. Penetration test: at least annually and after every significant architectural change. Before go-live: a full vulnerability scan with no open Critical / High vulnerability. Critical patches: in SaaS deployed within 7 days (internet-facing) / 30 days (internal); in on-prem per the agreed SLA. Pen-test summaries are provided under NDA.
What happens to our data after the contract ends?
In SaaS: before termination we let you export all customer data. After termination, data is removed from production; backups are rotated and deleted per the retention policy. On request we provide confirmation of data destruction. In on-prem: the data stays with you. Upon license termination we agree on migration or uninstallation.
What is your notification deadline in the event of a security incident?
In the SaaS models, in compliance with GDPR Art. 33 and the DPA, we notify the customer without undue delay, no later than 72 hours after confirmation of an incident endangering personal data. In the on-prem models, ATLAS GROUP notifies customers of security vulnerabilities discovered in the delivered software — incident response for operations is in the customer’s hands.
How long do you keep logs and who has access to them?
Standard logs: at least 6 months or as required by the customer. Logs involving the processing of personal data: 3 years (or as agreed). Access: only authorized staff on a need-to-know basis. In on-prem: logs stay with you, exportable to your SIEM (Splunk, Elastic, Sentinel, etc.).
When should you choose which deployment model?
Cloud (SaaS) — most law firms and companies. A fast start, no internal infrastructure, ATLAS GROUP manages everything. Private cloud — regulated entities (banks, insurers, healthcare). A dedicated instance, but still managed by ATLAS GROUP. On-premise — customers with an internal policy or a regulatory requirement for residency of the application and stored data in their own infrastructure. The AI path still goes through OpenAI; in all models you can optionally enable EU data residency.
Documents on request
Under NDA, within two business days
For your vendor security review we will prepare the documents listed below. Send your request to klientske.centrum@atlasgroup.cz.
- ISO / IEC 27001 certificate
- ISO / IEC 27017 certificate
- ISO / IEC 27018 certificate
- DPA — Data Processing Agreement (template, Art. 28 GDPR)
- NDA — template for vendor review
- Penetration test — summary (past year)
- BC / DR plan — summary (BIA, RTO / RPO)
- List of subprocessors — updated quarterly
- Shared responsibility matrix — for the individual deployment models