The whole environment.
One considered response.
Connect authorised telemetry, preserve useful context and investigate relationships—then act only within defined authority.
Keep the context.
Control the volume.
Bounded queues, aggregation, prioritised batches and gradual recovery are design requirements. They are not a claim of zero impact or lossless collection under every condition.
- 01CollectAuthorised sources
- 02NormaliseUseful shared context
- 03AggregateRetain counts and evidence
- 04PrioritiseBound every queue
- 05DeliverBudgeted, gradual recovery
Map relationships, not just inventory.
OmniGraph is the planned relationship model for accounts, devices, services, data stores and AI agents. A relationship must retain its source, confidence and last-observed time. “Can access” is different from “did access”.
Collection starts with explicitly authorised sources. Disconnected accounts, unsupported systems and expired permissions remain visible as coverage gaps, rather than becoming a misleading green status.
One evidence trail. Multiple possible explanations.
The investigation layer is designed to combine policy findings and related observations into a case. Model-generated text should point back to evidence and distinguish observation, interpretation and uncertainty.
Approved migrations, new devices and unusual working hours can be legitimate. The demo includes an authorised migration so you can see an investigation resolve without containment.
Authority is a separate system.
Read-only collection does not create write access. A response requires a supported connector or gateway, an authorised identity, a matching policy and any required human approval.
Guardian’s proposed modes are Observe, Assist and scoped Pre-approved Response. High-impact decisions should require explicit approval. Every proposal needs a target, scope, evidence, business impact, expiry where supported and a recovery explanation.
An investigator must never be able to expand its own permissions. OmniGuard does not retaliate against external systems.
Control points determine what can be prevented.
An AI tool gateway can evaluate requests routed through it. It cannot govern a separate credential or network path that bypasses it. Endpoint actions similarly depend on the capabilities of the installed or integrated agent.
Integration design must document bypass routes, lost connections, queue delays and failure behaviour. A successful API response is not enough: the platform needs to verify the resulting state.
Plan beyond the alert.
Exposure modelling and recovery orchestration are roadmap capabilities. The intended role is to connect known permissions, dependencies and existing recovery systems, not to invent complete attack predictions or replace backup storage.
A backup existing, a backup passing integrity checks and a restoration being successfully tested are different facts. Proposed recovery changes need service-owner review, an appropriate isolated test and a record of what actually recovered.
Evaluate the architecture before connecting.
The technical brief describes the intended data flow, enforcement boundaries, model choices, required permissions and open engineering decisions. No working production connector is advertised on this site.
Will it replace an existing security team?
Can it operate with a local model?
Will every company system work immediately?
Less noise.
More understanding.
Explore the idea. Challenge the assumptions.
Help shape what comes next.