The AI Governance Plane

Enterprise AI,
without the trade-off.

Bastion governs AI at the moment of inference — it classifies every prompt by data sensitivity, keeps regulated work inside your boundary, routes the rest to the best model for the task, and proves every decision. Cost‑effective, secure, compliance‑ready, auditable — by design.

Runs entirely inside your boundary · drop‑in OpenAI‑compatible · deploys as one container

0 regulated prompts
left the boundary across an adversarial red‑team corpus — enforced by design, fail‑closed.
97.6%
de‑identify recall on a held‑out corpus, with 0% benign over‑block (measured, in‑boundary).
~27 ms
median governance overhead per request — before the model is even called; a fraction of inference time.
Built for the inference path — on the infrastructure you already run: Nutanix NKP GPT‑in‑a‑box NVIDIA NIM vLLM / Ollama Any OpenAI‑compatible fleet
The gap

Inference is the open flank.

Enterprises spent a decade classifying and labeling data at rest and in motion. That investment stops the instant a model reads the data — and whoever governs that moment owns the control point for enterprise AI.

The regulated core hasn't moved

Health, finance, government, defense — the highest‑value AI workloads are blocked on evidence, not model quality. Nobody can prove what a model was allowed to see, request by request.

RAG leaks past your labels

A retrieval‑augmented prompt can pull a labeled document from a store and carry its contents to whichever model answers. Your existing DLP and entitlements never see that hop.

“Sovereign or frontier” is a false choice

Teams are forced to pick the best model or control of their data. Removing that trade‑off is the difference between AI that ships and AI that stalls in review.

Built for regulated AI teams: Healthcare Financial services Government & defense — anyone who has to prove what a model was allowed to see.
The product

One layer, four jobs.

A single control point between your applications and every model they use — classify, enforce, route, and prove, all inside your boundary.

01

Classify

Every prompt scored by data sensitivity — deterministic detectors plus named‑entity recognition running in your own boundary. Session‑aware, raise‑only.

02

Enforce

Regulated work is hard‑blocked to your boundary — a policy pre‑filter, fail‑closed. De‑identify‑then‑escalate is a standing rule: detected PII/PHI crosses only de‑identified.

03

Route

The cheapest model that clears a competency bar for the task — local‑first, cloud only when policy permits. The best answer per prompt, biased onto your own compute.

04

Prove

Every decision lands on a tamper‑evident, hash‑chained audit — host‑forgery‑resistant when the key is held off the machine. Customer‑held evidence, not log lines.

Sensitive work still gets the best model

Identifiers are tokenized inside the boundary, only the safe residue is sent, and the answer is rehydrated locally — so a frontier model does the work while the detected identifiers stay in‑boundary. Turns “we can’t use AI on that data” into a governed yes.

Content‑free observability

Governance events flow to your existing SIEM and OTLP collector as labels and counts — never prompt text, PII, or tokens. Prove what an observer can’t measure: that sensitive work stayed in‑boundary.

How it works

The order of the stages is the design.

Policy runs as a pre‑filter, fixing the candidate models before any scoring — so no later stage can reopen a boundary policy has closed.

1

Classify

Score the prompt by data sensitivity, in‑boundary. Once a session is flagged regulated, that floor only ratchets up.

2

Policy pre‑filter

Regulated is hard‑blocked to the boundary, fail‑closed. The permitted model pool is fixed here — before anything is scored.

3

De‑identify where permitted

Tokenize detected PII/PHI in‑boundary; send only the safe residue; refuse egress outright if detection degrades.

4

Route & execute

Pick the cheapest capable model from the permitted pool — local‑first. The routing decision is made in‑boundary, so it sends nothing anywhere.

5

Prove

Rehydrate the answer locally and write the decision to the tamper‑evident chain — labels and counts only.

# Point any OpenAI-compatible client at Bastion.
# Same wire format — governance happens in between.

from openai import OpenAI

client = OpenAI(
    base_url="https://app.bastionplane.ai/v1",
    api_key="$BASTION_TOKEN",
)

resp = client.chat.completions.create(
    model="auto",   # Bastion picks the best permitted model
    messages=[{"role": "user",
               "content": prompt}],
)

# → regulated? served in-boundary.
# → the rest? best model for the task.
# → every decision on a tamper-evident audit.
Enterprise‑ready

Tested. Integrable. Secure.

Not a demo — a product engineered to run in production, and to be verified on your own hardware.

Tested

  • Unit, integration & end‑to‑end HTTP suites against the real server
  • 0 leaks across our adversarial red‑team corpus (measured in‑house), plus prompt‑injection probes
  • Concurrency, fault‑injection & live‑model runs — every proof re‑runs on your hardware

Integrable

  • Drop‑in OpenAI‑compatible endpoint — existing apps work unchanged
  • Governance + operations API, governed MCP tools, and a web console
  • Enterprise auth built in — SSO (OIDC · SAML · JWT), deny‑by‑default RBAC, tenant isolation

Secure

  • Fail‑closed by design; deny‑by‑default on every model call; every decision on a tamper‑evident chain
  • Automated security assessment (SAST + dependency/container CVE + full active, authenticated DAST) — zero active‑attack‑class findings (the one high was a demo‑only false positive) as of our latest CI run (in‑house) — and continuous in CI
  • An independent third‑party penetration test and SOC 2 are the scoped next step, run alongside your pilot
Where it fits

A capability‑layer map — not a scorecard.

Cost/quality routers optimize model selection. AI gateways centralize access. DLP tools classify content. Bastion owns the decision above them — what the data is, and whether it may cross.

CapabilityBastionCost/quality routersAI gatewaysDLP tools
Per‑prompt data‑sensitivity classification
Regulated hard‑block to the boundary, fail‑closed
De‑identify‑then‑escalate as a routing rule
Cost‑aware routing on measured competency (in‑house)
Routing brain that never egresses the prompt
Tamper‑evident audit, host‑forgery‑resistant (external key)
Runs entirely inside your boundary

✓ core · ◐ partial / varies by product · ✕ not their focus. Capability presence from public product categories — not a benchmark. Every Bastion row is verifiable in the delivered package. As of 2026.

On your estate

Local by default. Cloud by exception. Every decision proven.

Bastion is a layer on your sovereign infrastructure, not an alternative to it. Every prompt it keeps in‑boundary is a token served on your own GPUs — a control that fills the estate, not just one that protects it.

  • One container above your in‑boundary model endpoints (NAI · GPT‑in‑a‑box · NIM · vLLM).
  • Register any OpenAI‑compatible fleet and govern it — register‑then‑route.
  • No data plane leaves your control — a cloud proxy structurally can’t offer this.

Bring your regulated AI workloads home.

Name two engineers and four weeks. Install the package on your hardware, run the proofs against your own data, and read the audit chain that comes out — none of it requires us in the room.

or book a briefing · read the security & architecture brief →

Self‑contained · reproducible on your hardware · nothing leaves your boundary