.png){kind=logo}
Use AI on the data you can't afford to leak
Your teams want AI on the data where it actually pays off — and exactly the data legal won't release to a model. enclaive makes that governable: pseudonymized prompts, confidential pipelines, enforced policy, sealed runtimes — putting boundary control around the data used in AI.
Works with hosted LLM providers and private models. One policy layer across all of them.
The executive mandate says scale AI. The security review says not with that data. The standoff has three predictable shapes.
The pilot that can't graduate
The use case works, the demo impressed, and production approval died in legal review because nobody can say what happens to customer data inside a third-party model. Every quarter it stays blocked, the projected productivity gain stays projected.
Shadow AI is already running
While official projects wait, employees paste contracts, code and customer details into whatever chatbot works. The data exposure your review board is trying to prevent is happening anyway — unmanaged, unlogged, and invisible until an incident makes it visible.
Tools that don't see the problem
DLP and proxy controls were built for files and URLs, not prompt-and-response flows. They can't distinguish a harmless query from one carrying PII in context, can't enforce per-model policy, and produce no audit trail a regulator would accept.
enclaive protects AI usage at three layers. Use one or stack all three — the policy model and audit trail are the same.
Sensitive entities — names, IDs, account numbers, diagnosis codes — are detected in prompts and retrieved context and replaced with tokens before the request leaves your boundary. Answers are re-identified only inside your perimeter, where policy allows. Requests route exclusively to approved endpoints, with allow/deny enforcement and full logging. The model stays useful; the data stays yours.
Documents are ingested, embedded and stored as vectors with confidentiality maintained through every step — including during processing. Your internal assistant can draw on contracts, case files and research data without the knowledge base itself becoming a new leak surface for infrastructure operators or co-tenants.
For the most sensitive cases, run models and inference inside confidential environments: memory encrypted during execution, workload identity attested, keys released only against proof of integrity. This is how AI runs on patient records and citizen data — the setting where “trust the provider” was never going to pass review.
The value driver's own test applies: what would you roll out in the next 90 days if data exposure weren't a concern?
The honest answer to shadow AI
Bans don't stop shadow AI; they just keep it invisible. The workable answer is a sanctioned path that's as easy as the chatbot tab — and safer by construction. Give employees approved AI with guardrails they don't have to think about, and the incentive to go around the rules disappears.
One gateway for every approved model, hosted or private — no per-tool review cycles
Policy enforced in the flow: what may be asked, with which data, to which model
Violations detected and blocked before exposure, not discovered after
A complete prompt-and-response audit trail when compliance asks who used what
Adoption metrics leadership can actually see: approved use cases in production, violations trending down
What AI and security owners ask
Does pseudonymization make the model's answers worse?
The design goal is exposure minimization with utility preserved: entities are replaced with consistent tokens so the model keeps the structure it needs to reason, and re-identification happens inside your boundary where approved. For most support, drafting and document tasks the effect is negligible; for edge cases, the sealed-runtime layer exists so the raw data never needs to leave your control at all.
Which LLM providers does this work with?
The firewall sits between your users and any endpoint — hosted providers such as Claude (Anthropic), Gemini (Google), OpenAI and Mistral, or private and self-hosted models you run yourself. Policy decides which models are approved for which data classes, and that list is yours to change without re-architecting anything.
We already have DLP and a secure web gateway. Why isn't that enough?
Those tools classify files and destinations. AI risk lives in prompt-and-response flows: context windows assembled from RAG, PII embedded mid-sentence, model-specific policies, and the audit question of who asked what. That's a different inspection point and a different policy model — which is why the firewall complements your DLP rather than replacing it.
Where does our data physically go?
Pseudonymization mapping tables and keys stay vaulted inside your boundary under your custody. The knowledge base and, on the sealed-runtime layer, the models themselves run in confidential environments on your choice of cloud — including European providers or on-prem.
How does this relate to GDPR and the EU AI Act?
Pseudonymization, purpose-bound access, audit trails and demonstrable technical controls map directly to GDPR obligations, and the logging and policy-enforcement layer produces the usage documentation AI governance frameworks increasingly expect. For the control-by-control view, see the compliance mappings — and validate specifics with your DPO, which we'll support with the technical facts.
Can we start small?
That's the intended path: most programs start with the firewall in front of one approved use case, add confidential RAG when an internal assistant needs the knowledge base, and reserve sealed runtimes for the workloads that justify them. One policy layer, three levels of protection, adopted in the order your risk demands.
Confidential AI is Garnet, standing on the Compute Fabric and the Trust platform.
PrivacyMode pseudonymization, confidential RAG and sealed model runtimes in one workspace.
AI workloads scheduled into confidential Kubernetes, including confidential GPU nodes.
Mapping tables and model keys stay inside your boundary, released on attestation.
Deploy and operate the AI stack across clouds from one console.
Bring the use case legal blocked
Tell us the AI project stuck in review — the data involved, the model you want, the objection on the table. We'll show you which layer unblocks it and what the audit trail looks like.
