Security & Compliance | Confident AI Docs

Overview

This page provides detailed security information for your security team’s review. It covers:

Network architecture and traffic flow
Integration points with your existing infrastructure
Authentication and access control
Data retention and storage
Encryption at rest and in transit
Logging and audit capabilities

All self-hosted deployments—regardless of cloud provider—follow these security principles. Provider-specific implementation details are covered in the respective deployment guides.

Networking

All traffic flows through your private network. There is no public internet ingress by default—users access Confident AI through your existing VPN, network peering, or private connectivity setup. The only outbound traffic is to LLM providers for evaluations (if using external models) and optional notification services.

Network isolation

Component	Network placement	Access
Load balancer	Private subnet (or public with restrictions)	Your corporate network only
Application services	Private subnet	No direct internet access
Database	Private/isolated subnet	Application services only
Object storage	Private endpoint	Application services only

Internet-facing option: For organizations that prefer public access (with authentication), the load balancer can be configured as internet-facing. This increases attack surface but simplifies access for distributed teams without VPN.

Integration points

The deployment supports several integration points with your existing infrastructure. All inbound connections come through your private network; outbound connections go through NAT/gateway with optional private endpoint support.

Integration	Direction	Protocol	Purpose
Corporate SSO	Inbound	SAML/OIDC	User authentication via your identity provider
LLM providers	Outbound	HTTPS	OpenAI, Azure OpenAI, Anthropic, or self-hosted models for evaluations
Private connectivity	Bidirectional	Private	Connect to your existing cloud infrastructure
Slack / Email	Outbound	HTTPS/SMTP	Alert notifications and evaluation reports
OTEL collectors	Inbound	gRPC/HTTP	Receive traces from your LLM applications
CI/CD pipelines	Inbound	HTTPS	Trigger evaluations from GitHub Actions, Jenkins, etc.

Private endpoints: All major cloud providers offer private endpoint services (AWS PrivateLink, Azure Private Link, GCP Private Service Connect) that allow integration traffic to stay within your cloud network, eliminating public internet exposure.

Outbound access to LLM providers is required for evaluations. If your organization restricts outbound traffic, ensure the following are allowlisted:

api.openai.com (OpenAI)
*.openai.azure.com (Azure OpenAI)
api.anthropic.com (Anthropic)

Alternatively, use self-hosted or on-premise LLM models that don’t require external network access.

Authentication

Confident AI does not maintain its own user database for self-hosted deployments. All user authentication is handled through your corporate identity provider via SAML or OIDC. This means user credentials never touch Confident AI systems—authentication tokens are validated against your IdP directly.

Method	Description
SSO (SAML/OIDC)	Primary authentication method via your identity provider (Okta, Azure AD, Google Workspace, etc.)
Google OAuth	Optional OAuth authentication for organizations using Google Workspace
API keys	Scoped per project for programmatic access (SDK, CI/CD, integrations)
Service accounts	Cloud-native IAM for service-to-service authentication (IRSA, Workload Identity, etc.)

Local accounts are not supported. All human users must authenticate via SSO. This ensures: - Centralized access control through your IdP - Automatic deprovisioning when employees leave - Consistent MFA enforcement - Audit trail in your identity system

API key security

API keys are used for programmatic access and have the following security properties:

Project-scoped: Each API key is tied to a specific project and cannot access other projects
Revocable: Keys can be revoked instantly from the dashboard or API
Audited: All API key usage is logged with request details
No user impersonation: API keys cannot perform user-level actions like changing settings

Rotate API keys regularly. We recommend rotating API keys every 90 days or immediately if you suspect compromise. The dashboard shows last-used timestamps to help identify unused keys.

Data storage and residency

All application data is stored within your cloud account in the region you choose. Confident AI (the company) has no access to your data—everything remains in your infrastructure.

Data type	Storage	Purpose
User data	PostgreSQL database	Projects, settings, user preferences
Traces & spans	PostgreSQL database	LLM application telemetry
Evaluation results	PostgreSQL database	Metric scores, test runs, experiments
Datasets	PostgreSQL database + Object storage	Golden datasets, uploaded files
Audit logs	Cloud logging service	Security and compliance auditing

Data residency: Your data never leaves the region you deploy to. This is critical for organizations with GDPR, data sovereignty, or regulatory requirements that mandate data stay within specific geographic boundaries.

Data retention

Retention periods are configurable per data type, allowing you to meet compliance requirements while managing storage costs. Data deletion is handled automatically based on your configured policies.

Data type	Default retention	Configurable
Traces & spans	90 days	Yes
Evaluation results	Indefinite	Yes
Uploaded files	30 days	Yes
Audit logs	Per cloud provider defaults	Yes
Application logs	14 days	Yes

Deletion is permanent. Once data exceeds its retention period, it’s automatically deleted and cannot be recovered. Ensure your retention settings align with your compliance and business requirements before deployment.

Retention periods are fully configurable. Contact your Confident AI representative to adjust based on your compliance requirements or modify settings in the application configuration.

Encryption

All data is encrypted both in transit and at rest using cloud-native encryption services. Encryption keys are managed within your cloud account—Confident AI does not have access to your encryption keys.

Layer	Method
In transit	TLS 1.2+ for all internal and external traffic
At rest (database)	Cloud-managed encryption (AWS KMS, Azure Key Vault, GCP Cloud KMS)
At rest (storage)	Cloud-managed encryption with bucket/container policies
Secrets	Cloud secrets manager with encryption (AWS Secrets Manager, Azure Key Vault, GCP Secret Manager)

Encryption key management

Option	Description	Recommended for
Cloud-managed keys	Default encryption using cloud provider’s managed keys	Most deployments
Customer-managed keys (CMK)	You create and control the encryption keys	Regulated industries, specific compliance requirements
BYOK (Bring Your Own Key)	Import keys from your HSM	Maximum control, on-premise HSM integration

Default is secure. Cloud-managed encryption meets most security requirements. Customer-managed keys add operational complexity but provide additional control for organizations that require it.

Logging and audit

Comprehensive logging is enabled by default for security monitoring and compliance. All logs are stored within your cloud account using native logging services.

Log type	Contents	Retention
Application logs	Service logs, errors, request traces	Configurable
Access logs	Load balancer access, API requests	Configurable
Audit trail	Cloud API calls, IAM actions, resource changes	Per compliance requirements
Database logs	Connection logs, slow queries (optional)	Configurable

What’s logged

User actions:

Login/logout events (via SSO)
Project creation, modification, deletion
Dataset uploads and modifications
Evaluation runs and results access
API key creation and usage

System events:

Service health and errors
Database connections and queries
External API calls (LLM providers)
Resource scaling events

PII in logs: Application logs may contain user-provided data (prompts, responses) depending on your tracing configuration. Review your logging settings and apply masking rules if needed to prevent sensitive data in logs.

Compliance considerations

Self-hosted Confident AI supports various compliance frameworks by design:

Framework	How self-hosting helps
GDPR	Data stays in your chosen EU region, you control retention and deletion
SOC 2	Deploy within your existing SOC 2 boundary, inherit your controls
HIPAA	Keep PHI in your compliant environment, sign BAAs with cloud provider
FedRAMP	Deploy in FedRAMP-authorized regions with your ATO
Data sovereignty	Data never leaves your geographic boundary

Your compliance, your controls. Self-hosting means Confident AI inherits your organization’s compliance posture. Work with your compliance team to ensure the deployment meets your specific requirements.

Security checklist for your team

Before approving deployment, your security team should verify:

Network architecture reviewed and approved
VPN/private connectivity plan confirmed
SSO integration method selected (SAML/OIDC)
Data retention policies defined
Encryption key management approach decided
Logging and monitoring integration planned
Outbound access rules for LLM providers approved
Incident response procedures updated to include Confident AI

Next steps

Once your security team has reviewed this information, proceed to the deployment guide for your chosen cloud provider:

AWS Deployment
Azure Deployment (coming soon)
GCP Deployment (coming soon)