A sovereign raw-data lake beneath a mature, accredited body of detection and compliance content, with machine learning and constrained generative AI that never leave your boundary. Software on certified hardware or in your accredited private cloud.
Build order is deliberate: data plane first, capability overlay second, machine learning third, generative AI last. Each layer depends on the one beneath it, and nothing is routed outside your boundary to be analysed.
Every capability reads the same complete data on infrastructure you control.
All telemetry lands once, raw, in open formats.
Logs, endpoint, network, identity, file activity, OT protocol events, IoT device events and AI traces land in one store on hardware you control, queryable in seconds from hot storage. A streaming backbone, a columnar analytical store and an open-table lake for retention. You set retention; the vendor does not meter it.
A mature body of content, running on complete data.
Real-time rule evaluation on the stream, with detection content expressed in the open Sigma standard alongside native rules. Correlation runs on entities, not on log lines, so the same actor touching an identity, a model and a data store is seen as one story.
Whether a control fired, not whether it is documented.
Control-level measurement of regulatory obligations from the same telemetry: government control baselines, operational-resilience rules and critical-infrastructure regimes. Continuous, evidenced, and mapped to the obligations a board is accountable for.
Ask any question of any data, then act inside your boundary.
Sub-second investigation over billions of events from one query surface. Response workflows run as data through an embedded orchestration engine; a human approves execution, and every action is recorded against the evidence that justified it.
Behavioural detection at millions of events per second.
Baselining, anomaly scoring and classification per entity and per protocol, against the analytical store. CPU-class work, trained in Withstand's build environment and shipped as signed artefacts through the same offline channel as software.
Explains, drafts and ranks, and is never the lead detection claim.
An open-weight model served on your own GPU tier, restricted to natural-language query, obligation-mapped incident narrative for boards, and response-playbook drafting. Nothing model-related leaves your environment, and your data never trains anyone's models.
Collection is where most platforms quietly narrow their claims. A sovereign environment contains all five, and the platforms you already run are sources too.
Endpoint process, file, registry and network events; server and application logs; identity, privileged access and MFA; network flow and DNS; firewall, proxy and email gateway; hypervisor and storage events. Collected by agent, event forwarding, syslog, agentless pulls and network sensor.
Asset inventory and protocol-level events from PLCs, RTUs, SCADA servers, HMIs and historians; control commands, set-point changes, firmware and logic changes. Observed passively on SPAN or TAP, crossing into the enterprise side through a data diode. Active scanning of OT assets is never performed.
Device discovery and fingerprinting; camera, sensor, building-management, access-control and medical-device traffic; broker and gateway events. Most devices cannot host an agent, so the value is behavioural baselining against each device's own normal.
Prompts, responses and tool calls through AI gateways; agent actions and privilege use; model registry and deployment events; vector store and retrieval access; guardrail outcomes. Captured as OpenTelemetry-compatible traces and joined to the identity that initiated them and the data they touched.
File and share access audit; database audit logs; classification and sensitivity labels; DLP and rights-management events; removable-media and cross-boundary transfer events. Lineage from source to model to output.
Cloud security platforms and specialist OT monitoring tools are consumed as sources, not replaced. Withstand is the sovereign layer beside what you already run, for the networks those platforms cannot serve.
| Test | What it requires | Why it matters |
|---|---|---|
| Operational | The platform runs inside your boundary; nothing phones home; updates arrive by controlled media. | Any external dependency is an unassessed path into the most sensitive system on the network. |
| Jurisdictional | The vendor is owned, controlled and answerable inside your jurisdiction. | Foreign ownership, foreign capital and foreign legal obligations survive on-premises deployment. Running the software yourself does not change who the vendor answers to. |
| Supply chain | Documented provenance of every component; imaging and hardening performed in-country; no excluded-vendor parts. | Accreditation attaches to specific bills of materials. Supply chain is now a primary attack surface. |
| AI data governance | Your data never leaves your environment and never trains anyone's models, the vendor's or a third party's. | Digital sovereignty has moved past privacy to control over national-interest data and AI training material. |
Operational sovereignty means you run the software. Jurisdictional sovereignty means the vendor answers to your law. Only the second is a matter of law rather than trust. Withstand is independently owned, carries no intelligence-community capital, is delivered and hardened in your jurisdiction, and is designed so that inference happens inside your accreditation boundary. It is built to clear all four tests.
Withstand will never lead with an 'AI-powered detection' claim. Its detection credibility comes from proven content augmented by classical machine learning; generative capability is applied only where it is the right tool. Models are open-weight with published provenance and a model bill of materials, signed alongside each release.
A signed image on certified hardware or in your accredited private cloud. Single site, clustered, hub and spoke, or one instance per classification domain. Nothing leaves the boundary.
Your own servers from Withstand's hardware compatibility list, imaged and enrolled with a signed operating-system image, measured boot and drive binding. For agencies bound to hardware panels and estates that already own the metal.
The same signed image as a virtual machine in your already-accredited private cloud. The trust posture is reduced and stated as such: the hypervisor is in the trust base.
A core instance plus edge collectors and passive sensors at remote sites and OT zones, joined by diodes or constrained links. Collectors hold local evidence if the link is lost and reconcile on reconnection.
One instance per classification domain, with one-way flows through cross-domain guards where policy permits. Updates, content and models arrive on signed media and are verified against the trust root before use.
Withstand is engineered on the assumption that it will be targeted. An attacker who reaches the monitoring layer can blind the organisation, alter evidence and hide.
A signed, immutable operating-system image on certified hardware, with measured boot from firmware to container images. The system refuses to run altered software and can prove what it is running.
Retained telemetry and audit logs are written to write-once storage tiers with cryptographic chaining, so an intruder who reaches the platform cannot silently alter history.
Services run isolated with per-service identities. Administrative access is role-based, multi-factor and fully audited. No shared credentials, no vendor back door.
FIPS-validated encryption at rest with keys bound to the TPM or an external HSM where mandated. Crypto-erase supports classified decommissioning.
The platform monitors itself with the same detection engine and raises alerts on its own integrity, configuration and access events to a separate, minimal channel.
Offline, signed backups of configuration, detection content and models, and documented bare-metal recovery to a known-good state without any external service.
The next-generation Withstand platform is being built with design partners whose constraints are the specification. If your networks cannot leave, you are who this is for.
Become a design partner