"> Security & Architecture — Governe by Aiebo

Security & architecture

We see what's in your prompts. We never see your prompts.

Both planes run single-tenant inside your environment. Governe inspects and governs every AI request locally, while administration, audit, and reporting remain under your control. Customer content is never sent to Aiebo.

Data handling

What our control plane sees — and what it never does.

Data	Reaches Governe control plane?	Notes
Prompts, completions, embeddings, documents	Never	Exist only inside your single-tenant runtime plane and the model providers you route to.
Classification findings (e.g. "38 PII fields detected, blocked")	Yes	The finding travels; the text that produced it does not. This is what makes reporting possible without exfiltration.
API keys & credentials	Hashes only	Raw keys never leave your environment; revocation propagates in ≤ 60 seconds.
Usage & cost telemetry	Aggregates, async	Buffered locally (durable queue), batched upstream; survives network loss.
Inventory metadata (app names, owners, models, vendors)	Yes	The minimum needed to attribute activity and report on it.
Policy definitions & audit events	Yes	Signed policy travels down; enforcement decisions are logged for evidence.
Training data for our models	Never	We do not train on customer data. Contractually.

Design principles

Built for the failure cases.

Inspection is local by construction

Classification models and detection rules run inside your single-tenant runtime. There is no code path that ships prompt or document content to our control plane — not for analysis, not for debugging, not for model improvement.

Runtime never calls home in the request path

Every enforcement decision is made locally from signed policy snapshots. Governe cloud latency — or an outage — can never slow or block your AI traffic.

Offline-first, fail-safe by policy

Intermittent or air-gapped networks enforce from cache with a configurable staleness window — fail-closed or allow-with-audit, your choice.

Bounded overspend via budget leasing

Budgets are enforced as leases, so even a disconnected runtime cannot exceed its granted allowance.

≤ 60-second revocation

Compromised keys and revoked agents propagate to every enforcement point within a configurable SLA of one minute.

Evidence by default

Every policy change, approval, and enforcement decision is recorded — versioned, backward compatible, ready for auditors.

Deployment

Run it where your risk profile says.

Recommended for most enterprises

Customer-deployed

Full data residency and control — Kubernetes, VM, or bare metal
Works offline from cached, signed policy
You operate the gateway in your VPC or data center
Local storage for audit buffer (SQLite/Badger)

Managed

Governe-operated · single-tenant

We operate the gateway for you — same architecture, same guarantees
Deployed in your isolated, single-tenant VPC
Encrypted storage, Kubernetes-managed
Same boundary: single-tenant, and your content never reaches our control plane

Why not the tools you already have?

Three categories. Three blind spots.

Each of these solves a real problem. None of them answers "what data went to which model, sent by whom, and was it worth it?"

DLP & CASB

They watch the network.

Built for files and web sessions, not prompts. They can tell you someone visited a chatbot — not that they pasted 1,240 customer records into it, or that an agent did it 400 times an hour.

No visibility inside the prompt

AI gateways & routers

They enforce traffic.

Routing, guardrails, and budgets — but only for the requests developers deliberately send through them. The SaaS copilot your marketing team turned on last Tuesday never touches it.

Blind to everything outside the gateway

GRC & governance platforms

They document policy.

Registries, frameworks, and attestations — a careful record of what AI should do, assembled from what people remembered to declare.

No evidence of what actually happened

Governe does discovery and enforcement — on the same record.It finds AI nobody registered, inspects what those requests actually carry, blocks what shouldn't leave, and prices the rest. Because it runs single-tenant in your network, seeing inside your prompts never means shipping them to us.

Straight answers

What a skeptical CISO asks next.

How can you classify our prompts without reading them?+

Because the part of Governe that reads them runs on your infrastructure, not ours. Every deployment is single-tenant inside your network. Inspection, PII and IP detection, redaction, and enforcement all happen there. What crosses to the control plane is the finding — "1,240 PII fields, blocked, Support team, ChatGPT" — never the text that produced it.

How do you see AI we don't route through you?+

Three ways, layered. Identity and SSO signals show which AI services people have authenticated to. Network egress and SaaS admin data show which AI vendors are receiving traffic. Expense and billing data catch the tools bought on a card. Anything you route through the gateway gets full content-level inspection; everything else still gets discovered, attributed, and priced — so nothing sits entirely outside the record.

Is this just another AI gateway?+

The gateway is one enforcement point, not the product. A gateway only governs what developers choose to send through it. Governe's job is the whole estate — including the SaaS copilots, embedded vendor AI, and personal accounts that will never touch your gateway — and then pricing all of it against what it returned. If you already run LiteLLM or similar, Governe sits above it rather than replacing it.

Does it slow down or break our AI traffic?+

Enforcement decisions are made locally from cached, signed policy — there is no call to Governe in the request path, so our latency or availability can never affect yours. You start in observe-only mode and see what would have been blocked before anything actually is.

We already have DLP. Doesn't that cover this?+

DLP was built for files moving across a network boundary. It can see that a browser session went to a chatbot; it generally can't see the 1,240 customer records inside the prompt body, can't tell an employee from an autonomous agent, and has no concept of which model or vendor is on the other end. It also has nothing to say about cost or return.

What about agents that act on their own?+

Agents are treated as first-class actors with their own identity, credentials, human owner, and approval record. You can see which agents can act autonomously, what data each has reached for, what it sent onward, and who signed off. An agent without a named owner is itself a finding.

What if our network is intermittent or air-gapped?+

Enforcement is offline-first. The runtime works from cached, signed policy with a configurable staleness window and fail-closed or allow-with-audit behaviour, your choice. Telemetry buffers locally in a durable queue and retries.

How fast do we see something real?+

The 30-minute assessment produces your first exposure map. Read-only discovery live in about a week; content inspection and first policies in roughly two; full estate coverage across teams and providers in four to six.

Compliance status — stated plainly

We publish our posture honestly rather than borrowing badges. Controls are designed against SOC 2 and ISO 27001; HIPAA-ready deployment patterns are supported. Certification status below is kept current — ask us for the detailed control mapping under NDA.

SOC 2 — aligned · audit in progress ISO 27001 — aligned HIPAA — deployment-ready No training on customer data — contractual

Security contact: security@aiebo.ai · We welcome coordinated disclosure.

Schedule an AI Assessment