Private development
The products consolidate working systems the founders have already built; Kordane is opening its first design-partner pilots.
Security & trust
Controls, threat model, and current stage
No badges, no absolutes. This page explains the mechanisms the platform is built on, each with its status, the risks the design reduces, the risks it does not claim to solve alone, and exactly where Kordane is today.
Trust architecture
Nine controls you can interrogate.
Each is a mechanism, not a promise: Kordane documents the mechanism, its current maturity, the trust boundary and the known limitations behind every control. What a reviewer can reconstruct after an execution is on theexecution evidence page; architecture and control documentation goes deeper during qualified pilot discussions.
Local-first transformation
Detection and tokenisation run inside your environment, before anything crosses.
Demonstrated
Model & tool allowlists
Only destinations you have approved are routable by policy.
Demonstrated
Policy decisions
Each crossing is evaluated against explicit, versioned policy, not convention.
Demonstrated
Identity & access context
Requests carry who is asking, from where, under which role.
Available in pilots
Encryption & isolation patterns
Designed for encrypted transport and storage, with per-environment isolation.
Available in pilots
Human approval gates
Consequential actions can require a named person to approve before execution.
Demonstrated
Audit & replay
A detailed audit trail records each decision; crossings can be reviewed after the fact.
Demonstrated
Evidence-backed outputs
Outputs can be linked to the sources that support them, or flagged when they are not.
Available in pilots
Data retention controls
You decide what the boundary keeps, for how long, and where.
Available in pilots
Threat model
Threat model
A security product should say what it is for. Boundary is a control layer for information crossings: it reduces specific risks and does not claim to solve others alone.
Boundary across the portfolio
- Cortex. Knowledge, document and speech inputs cross through approved policies.
- Forge. Datasets, models and runtime decisions are evaluated and recorded.
- Axon. Conversation content and automated actions are subject to channel and policy controls.
Designed to reduce
- Sensitive-value egress. Restricted identifiers, references, and codewords leaving with an AI request.
- Unapproved model or tool use. Requests reaching models or tools outside your allowlist.
- Unaudited AI actions. Crossings and actions that leave no reviewable record.
- Unattributable outputs. Generated claims with no link to the evidence behind them.
Not solved by the boundary alone
- Endpoint compromise. A compromised workstation or host sits before the boundary. Endpoint and OS security remain yours.
- Insider misuse outside the boundary. A person or administrator with legitimate access acting outside monitored paths.
- Policy misconfiguration. The Boundary enforces the rules configured for it. Incorrect or overly permissive policies can still allow unsafe behaviour and need operational review.
- Prompt injection and unsafe tool behaviour. Input controls, route restrictions, approval gates, and audit trails reduce exposure. No boundary eliminates every adversarial prompt or downstream tool-execution risk.
- Model correctness. The boundary governs what a model sees and where it runs, not whether its answer is right.
- Upstream model and infrastructure integrity. The supply chain of the models and provider infrastructure you choose to run is a shared responsibility.
Portfolio-specific limitations
- Model evaluation is task-specific. Benchmarks measure the evaluated scope only; they do not establish general performance.
- Small-model success is not general capability. A small model clearing one task threshold implies nothing about other tasks.
- Benchmark leakage and dataset defects. Training and evaluation can inherit or amplify dataset defects and leakage; data quality remains a shared responsibility.
- Local execution is not security. Running locally reduces egress exposure but does not by itself make a system secure.
- Fallback routes can increase exposure. Escalating to a larger external model widens the crossing; fallback paths need the same policy scrutiny as primary routes.
- Channel connectors depend on platforms. Axon conversation connectors depend on external platform controls and availability.
- Human agents are a boundary. People handling escalated conversations remain part of the trust boundary.
Maturity
Current stage
Where Kordane is today.
Synthetic public demonstration
The demonstration on this site runs deterministic, local logic on fictional data.
Production controls scoped per pilot
Deployment hardening, key management, and retention are defined during pilot design.
Security documentation on request
Architecture and control documentation is shared during qualified pilot discussions.