Security
Security at Kauli starts with a simple principle: we don't touch what isn't ours. We build AI products on top of the infrastructure and models you trust, with encryption and isolation at every layer we control.
"We protect what we control. We're transparent about what we don't."
Most of our products leverage third-party LLM providers chosen by you or recommended by us. That means your inference data flows through providers like OpenAI, Anthropic, or Google — governed by their security practices and your agreement with them. What we control — your application data, your configurations, your workflows — is encrypted, isolated, and yours.
Architecture
Stored by Kauli. Encrypted. Isolated.
Your project data, configurations, documents, and workflow state are stored in our secure cloud infrastructure — encrypted at rest (AES-256) and in transit (TLS 1.3), with each tenant’s data held in its own dedicated database rather than a shared table with a filter. Access is least-privilege and logged.
Managed by Kauli, governed end to end.
The platform’s AI runs on frontier models under Kauli’s own governed accounts — nothing for you to wire up — with each task routed to a model suited to it. Enterprise deployments can substitute self-hosted models or, as a premium configuration, your own provider keys. Kauli never trains models on your data. We send only the minimum context each request needs and never cache model responses beyond your session. Inference data is governed by your agreement with that provider.
Private TEE for clients who need it.
For organizations that require the highest level of data protection, our enterprise track can provide deployment inside private Trusted Execution Environments (TEEs). Your data and AI inference are processed inside isolated, encrypted enclaves, with inference running within the enclave rather than through a third-party AI API — so no one outside it can access your data or your prompts: not Kauli, not the infrastructure provider, not an AI model provider.
These are the security measures we apply to the infrastructure, application layer, and data storage under our direct management.
All application data is encrypted using AES-256 at rest and TLS 1.3 in transit. Database connections, API calls, and internal services all communicate over encrypted channels.
Our isolation model gives each tenant a dedicated database, backed by row-level security inside shared control surfaces. Cross-tenant access is architecturally prevented — not just policy-prohibited — and we test that boundary adversarially.
Role-based access with least-privilege principles. Internal access to production data is logged and requires explicit authorization. We keep our team's access surface as small as possible.
Your data is yours. We don't use it to train models, sell to third parties, or retain beyond what's needed to deliver the service. You can request deletion at any time.
Building with AI introduces unique risks. We design for them explicitly.
Every Kauli agent operates as a scoped principal under the permission envelope of the person it works for. No permission, no access — agents fail closed, sensitive domains carry a hard floor no agent crosses, and every agent action lands in an audit trail.
Our products include guardrails against prompt injection, data extraction attempts, and adversarial inputs. We treat AI security as a distinct discipline with its own threat models — and we run an in-house red team against our own releases.
We send only the data necessary for each inference request. Sensitive fields can be redacted or anonymized before reaching the LLM provider, depending on your configuration and security tier.
Kauli never fine-tunes or trains models on your data. The one exception is a fine-tune you explicitly commission — trained on your data, for your use, and it remains yours. When you connect a third-party model, we help you configure the connection to opt out of the provider’s training programs wherever the provider supports it.
Autonomy is earned, not switched on. Every agent starts supervised and charges a trust battery as it proves reliable, and outward-facing or irreversible actions sit behind human-in-the-loop approval gates, tiered by risk — routine work flows, consequential work waits for a person.
Agents learn from your operation under governance, not by osmosis. What enters your firm’s knowledge base is gated for relevance and appropriateness, structural changes require human approval, and every entry carries provenance — so a single bad input can’t quietly poison what your agents know.
Agents do their work inside contained environments with bounded tools and controlled egress — not loose on your systems. Security agents review agent behaviour continuously, and an administrator can revoke any agent’s access instantly.
We're transparent about which LLM providers are used in each product and deployment. You'll always know where your inference data is going and under what terms.
We design with compliance in mind from the start, not as an afterthought.
Our systems are built to meet GDPR and CCPA requirements at the architecture level. As we grow, we'll pursue additional certifications and update this page accordingly. If your organization has specific compliance needs, we're happy to discuss how our approach aligns.
If you've discovered a vulnerability in any Kauli system, we want to hear about it. Please report security issues to [email protected]. We'll acknowledge your report within 48 hours and work with you to resolve it.
We're happy to walk through our architecture, discuss your compliance requirements, or explain how the TEE option works for high-security deployments.
Get in Touch →