Security & data protection

Control access.
Minimize retained data.

Kyberos is being built around a simple principle: operating an AI application should not require keeping a second copy of every conversation.

Private development · This page describes the initial gateway foundation and the planned security model. It does not represent a security certification or an independent audit.

Explore Data Protection & Guardrails →
Current gateway data flow

Understand the path of a request.

  1. 01 / Application

    Your request

    Your application sends the prompt to Kyberos using a Kyberos API key.

  2. 02 / Kyberos

    Authenticate & forward

    The gateway processes the request in memory and forwards it using the configured provider credential.

  3. 03 / Provider

    Generate & return

    The provider processes the request. The response passes through Kyberos back to your application, including streaming.

What belongs in telemetry

The foundation exposes a metadata-only telemetry hook. Operational fields such as timing, status and request identifiers can support diagnostics; usage and cost analytics are being developed separately.

What stays out of content logs

Prompt bodies, response bodies and credentials are not intended for persistent gateway telemetry. This does not mean Kyberos never processes content: forwarding a request requires handling it in memory.

Kyberos Data Protection & Guardrails

Decide what reaches the model.

Our planned protection layer would inspect model traffic, apply your project rules and replace or block sensitive content before it reaches a provider.

Planned Kyberos capability · Not available in the initial gateway. Detection accuracy and provider compatibility have not yet been validated.

  1. 01 / AccessAuthenticate

    Resolve the project and its protection policy.

  2. 02 / ContentInspect & protect

    Detect findings, then allow, mask or block.

  3. 03 / InferenceForward approved content

    Only requests that pass the policy proceed to the configured provider.

  4. 04 / ResponseCheck the return path

    Apply output rules before returning the result.

Application input

Draft a reply to alex@example.com about order 4821.

Model input after masking

Draft a reply to [EMAIL_1] about order 4821.

Illustrative email-only policy. Other fields, including order IDs, need their own classification rules. Masking selected fields does not make the entire request anonymous.

Detect sensitive data

Identify structured personal data such as email addresses and payment card numbers, plus recognizable API tokens and private keys. Names and free-form addresses need additional detection and evaluation.

Apply project rules

Choose whether to allow, mask or block a finding. Keep replacements consistent within a session and stop a request if a required protection check fails.

Protect agent context

With a runtime adapter or MCP proxy, filter tool results before they enter model context. Limit a CRM response to the fields the task actually needs.

Audit decisions, not secrets

Record finding categories, counts, actions and policy versions. Keep original values and reversible mappings out of telemetry.

Start with a testable scope.

The first proposed prototype covers non-streaming text requests and deterministic rules for structured data and recognizable secrets. Streaming, attachments, images and tool-call transformations require separate support and tests.

Restore only with permission.

A later phase could restore masked values for an authorized recipient using encrypted mappings scoped to a tenant and session, with short retention. Models and subagents would not receive unrestricted access to those mappings.

One layer, replaceable detectors.

Kyberos would own policy enforcement and the integration interface. Detection engines could be local or explicitly approved external services, selected according to deployment requirements.

Clear protection boundaries.

Server-side inspection processes original content in memory. Detection can miss sensitive data or block harmless text. These guardrails are not a guarantee against prompt injection, a sandbox for agents, or a compliance certification. Direct tool access needs its own enforcement.

Discuss data protection requirements
Gateway foundation

Protection starts in the request path.

The initial gateway establishes these controls. Deployment configuration and verification remain part of making a system ready for production.

Separate credentials

Applications authenticate with a Kyberos key. The gateway uses a separate upstream credential, reducing the number of applications that need direct access to a provider secret.

Explicit header forwarding

An allowlist controls which request headers reach the provider. Kyberos authentication is replaced by the upstream credential; application cookies and internal Kyberos headers are not forwarded.

Bounded gateway traffic

Request-size limits, concurrency limits and upstream timeouts bound gateway resource use. Streaming connections are closed when the client disconnects; this is not a guarantee that the provider has stopped all processing or billing.

Restricted container runtime

The initial container setup uses a non-root user, a read-only filesystem, dropped Linux capabilities and no-new-privileges. The local Compose service binds to loopback; production ingress and transport security depend on deployment.

Planned controls

Built toward stronger project boundaries.

These are roadmap commitments in the design, not controls to assume are available in the current gateway.

Planned

Project isolation

Derive organization and project identity from the authenticated key, enforce authorization on every resource, and add database row-level security as a second layer.

Planned

Key lifecycle & secret management

Store only hashes of Kyberos keys, add expiration, revocation and rotation, and resolve provider secrets through a secret manager using workload identity.

Planned

Usage policy & audit history

Add per-project model allowlists, rate limits, budgets and a kill switch. Record administrative changes without logging secrets or prompt content.

Planned

Retention & administrative access

Add explicit metadata retention and deletion controls, redact sensitive metadata, and separate human administration through SSO/MFA and role-based permissions.

Discuss your data requirements before adopting.

Share your requirements for hosting region, provider access, retained metadata and administrative controls when discussing beta access. Availability needs to be confirmed for your deployment.

Discuss beta requirements Read the architecture