Safeguard
Solution · Sovereign Deployment

Air-gapped. Full Griffin lineup. No internet egress.

Safeguard runs in sovereign and classified environments on customer-owned GPU, with no internet egress, full audit log export, and the entire model family — through Griffin Zero (671B-MoE) — installed on-prem. Built for regulated public-sector workloads and defence supply chains.

◈ the standing query — evidence answers on demand, not once a year
STQC
Ready
FedRAMP
HIGH Ready
CMMC
L3 Aligned
0
Egress
What sovereign means here

Three guarantees, no asterisks.

Sovereign is not a marketing tier. It is a deployment topology with measurable properties — and we ship the evidence.

Physically isolated control plane

Control plane and inference cluster live on customer-owned hardware. No shared tenant boundary, no shared key material, no shared logs.

Customer-controlled keys + audit

Bring-your-own KMS, hardware-rooted signing, and a full audit log export every action emits — in JSON and CycloneDX, signed.

Full model lineup on-prem

Lion at the edge, Eagle in the cluster, Griffin Lite through Griffin Zero (671B-MoE) all run on-prem with no calls to Safeguard's cloud.

Deployment shapes

Same brain, four isolation levels.

Pick the shape that matches your data-residency and trust boundary. The lineup stays consistent end-to-end.

Shared cloud · Multi-tenant

Multi-tenant inference, fastest onboarding, per-tenant isolation at the prompt and KV cache layer.

Dedicated cluster · Single-tenant

Single-tenant inference on isolated VPC hardware. No cross-tenant traffic, deterministic latency, SHA-pinned weight attestation.

VPC-isolated · Customer VPC

Customer-controlled VPC, bring-your-own-key encryption, no cross-tenant traffic. Inference plane sits inside your network perimeter.

Sovereign / air-gapped · On-prem GPU

No internet egress, on-prem GPU, full audit-log export. Suits regulated, classified, and sovereign workloads.

Compliance alignment

Frameworks the deployment is designed to satisfy.

Pre-mapped control narratives, evidence packages on demand, and exports in the formats your auditor already accepts.

SOC 2 Type II
ISO/IEC 27001:2022
FedRAMP HIGH
CMMC Level 3
NIST SP 800-161
EO 14028
NIS2
DORA
STQC
Customer-specified regulatory frameworks
Operations posture

How it runs inside the wire.

Posture, in writing

  1. 01

    Signed model weights

    Model weights signed and verified at install with SHA-pinned attestation.

  2. 02

    Per-tenant inference isolation

    Per-tenant inference isolation — no cross-tenant prompt or KV cache reuse.

  3. 03

    Signed audit log export

    Full audit log export in JSON and CycloneDX, signed at emission.

  4. 04

    On-prem GPU sizing

    On-prem GPU sizing scales from 2x H100 (Growth) to 22x H100 multi-AZ (Mature).

  5. 05

    Customer-controlled break-glass

    Customer-controlled break-glass workflow with hardware-rooted approval.

Where the risk lives today

Four surfaces sovereign workloads have to defend.

No internet egress

Vulnerability feeds, model updates, telemetry — all need approved one-way conduits. The platform has to ship signed, verified, offline bundles, not assume a callback URL.

Customer-controlled key material

Model weights, audit logs, every piece of evidence must stay inside the boundary. Bring-your-own KMS, hardware-rooted signing, and audit exports your team alone can read.

Sovereign accreditation

STQC, FedRAMP HIGH-ready, CMMC L3 require explicit architectural decisions — control plane topology, key custody, evidence retention — not a retrofit at audit time.

Full model lineup parity

Sovereign customers should not get a watered-down lineup. Griffin Zero (671B-MoE) runs on-prem on customer hardware, with the same SHA-pinned attestation as shared cloud.

Current threat landscape

What the threat model actually contains.

Nation-state-class adversary

Adversarial robustness is a release gate, not an aspirational metric. Every build is regression-tested against red-team prompts.

Guardian runtime guardrails

Supply-chain compromise of training data

Corpus provenance and signed weights verified at install. Every artefact carries a chain of custody.

Signed SBOM + attestation

Insider risk

Capability scoping on agent tool calls, audit-log signing, hardware-rooted break-glass approval.

Capability scoping + audit

Coordinated disclosure inside the boundary

Internal-only research/disclosure mailbox; no external CVE channel touches classified findings.

On-prem disclosure workflow

Regulator audit packets on demand

Evidence is a query against signed history — not a project assembled fresh for each review.

Compliance evidence pipeline
Quantified benefits

Quantified benefits for sovereign deployments.

Measured on an air-gapped programme of record — same accreditation bar, dramatically less manual evidence work.

Metric
Before Safeguard
With Safeguard
Audit packet generation
3 weeks
1 hour
Critical-CVE remediation (air-gapped)
6-8 weeks
5 days
Model lineup parity with shared cloud
60%
100%
Evidence retention + export
Ad-hoc
Automated, signed
Adversarial regression detection
Monthly review
Per-build
Tool consolidation across SCIF
7+ tools
1
STQC / FedRAMP / CMMC evidence
4 separate packets
1 unified packet

On your hardware. Under your keys.

Talk to the team about on-prem GPU sizing, signed offline bundles, and the evidence packages your ATO or sovereign-readiness review needs.

Self-healing security runs on Safeguard.

Your first fix PR is minutes away.

No sales call required, even your agent can complete the purchase over MCP.