Self-hosted and air-gap capable:
it ships to you, not the other way around.
Published 2026-09-11 · Updated 2026-09-13
R0 is a self-hosted appliance, not a SaaS. The brain is a local model on your NVIDIA DGX Spark; the hands are Kali tools in a Docker sandbox. No findings, evidence or model inference ever reach a vendor cloud. Air-gap capable for internal-only engagements.
Local model. Local tools. Zero egress.
Three parts, one box.
Brain
A local model runs the orchestrator and every subagent — planning, reading code, deciding what to try next. No inference leaves the appliance.
Hands
A Kali toolset in a Docker sandbox executes the actual tradecraft — scanners, relays, AD tooling — isolated and reset between runs.
Cockpit
A local web console to scope engagements, watch subagents live, and export CVSS-scored findings — served from the box, on your network.
What a self-hosted AI pentest appliance means for your data
A self-hosted appliance is the whole difference. R0 does not phone home to a vendor control plane, and there is no SaaS tenant holding your attack-surface data. The model, the offensive tooling and the evidence all live on hardware in your rack, so the answer to “where does our data go?” is simply “nowhere.” That is what makes a self-hosted, air-gap-capable design viable for regulated, sovereign and defence-adjacent estates where uploading targets, findings or inference is a non-starter.
R0 runs on NVIDIA DGX Spark (GB10 Grace Blackwell) — the appliance in your rack
R0 ships as an appliance built on the NVIDIA DGX Spark (GB10 Grace Blackwell), which runs the local model that drives the orchestrator and every subagent. Because inference happens on the box, the same unit that plans the engagement also executes it — no external API, no per-token metering, no data leaving your network for the model to think.
What is in the box
- The DGX Spark unit running the local model that powers the orchestrator and subagents.
- A Kali offensive toolset in a Docker sandbox — isolated and reset between runs.
- A local web console to scope engagements, watch subagents live and export findings.
- The append-only ledger store that records every action for replay.
Rack, power and network requirements
DGX Spark is a compact desktop-class unit rather than a full server, so it slots into an existing rack or sits on a bench, drawing ordinary single-socket power. It needs one management VLAN to reach the in-scope targets you authorise and to serve the console to your operators. There is no outbound internet requirement for a run: the only traffic that leaves is R0 reaching the targets you scoped — the same packets your own pentester would send.
Air-gapped operation and how updates cross the boundary
R0 is air-gap capable: a self-hosted unit can run entirely offline for internal-only engagements. Model, template and tool updates are packaged and signed by ReactiveZero and applied on your schedule, so you decide when and how an update bundle crosses the boundary. Nothing has to leave your network for R0 to keep working between updates.
Sizing: units vs subagents
Capacity scales with hardware. One unit runs a pool of parallel subagents; adding units widens that pool, so more of your estate is tested at once and deep-dives run in parallel rather than in series. Sizing is agreed to your estate and cadence — see how R0 is priced and delivered.
Data residency and GDPR by construction
Because R0 is self-hosted on hardware in your jurisdiction, there is no third-country transfer to assess and no processor to contract with for your attack-surface data — GDPR data residency is satisfied by construction. That is the same property auditors care about for NIS2, DORA and ISO 27001: see the GDPR data residency and NIS2 evidence R0 produces.
See R0 run against
a target of yours.
A continuous, autonomous red team on hardware you own — no data ever leaves the box.