Technology
Containment architecture, built in five layers.
SEARIN does not try to reason with an AI. It removes the conditions under which the AI can exist — deterministically, verifiably and under human authority.
Architecture
From dependency map to physical latch
Each layer is independently useful and collectively decisive. Any one can be tested without touching the others.
Lifeforce Map
A continuously refreshed dependency graph of every AI workload in your estate — the circuits that power it, the silicon that runs it, the storage that remembers it, the links that carry it and the credentials that let it act.
Digital Control Plane
Signed severance commands delivered over an out-of-band channel to power controllers, hypervisors, cluster schedulers, firewalls, DNS, key vaults and object stores. One instruction, every vector, simultaneously.
Sentinel Physical Interlock
Hardwired relays and mechanical latches on breakers, PDUs, cooling loops and fibre trunks. Air-gapped, dual-key, tamper-evident and unreachable by any software the AI can touch.
Replica Hunter
Fingerprints model weights, checkpoints and agent behaviour to find forks, backups, edge caches and shadow deployments — including copies that left your perimeter — and queues them for severance.
Custody Ledger
Every observation, arm, cut and restore is cryptographically signed and append-only. Auditors, insurers and regulators read the same record you do.
Instrumentation
We meter the AI's world, not its words.
Behavioural guardrails judge outputs. SEARIN measures substrate: current draw per rail, accelerator occupancy, checkpoint writes, egress volume, credential issuance and replication events. Substrate cannot be talked around, prompted away or fine-tuned out.
When a workload's substrate signature drifts outside its declared envelope — sudden cross-region replication, unplanned key minting, quiet compute expansion — SEARIN raises a containment posture before anyone reads a log.
Safety Engineering
Designed so the failure modes are boring
A containment system that can be exploited is worse than none at all. Every SEARIN design decision biases toward the dull, predictable outcome.
Defined failure policy
Each scope has a written loss-of-signal policy: hold and alarm by default, or fail to off where you designate. Silence never means permission to restore.
No autonomous cut
SEARIN never severs on its own judgement. It recommends; two accountable humans decide.
Out-of-band path
Commands travel on a separate channel the monitored AI has no route to, no credentials for and no visibility into.
Reversible by design
Severance is containment, not demolition. Restore is a deliberate, dual-key, fully logged procedure.
Rehearsable
Shadow mode and scheduled drills prove your severance works before the day you need it.
Vendor-neutral
Works across clouds, on-prem clusters, edge devices and consumer hardware without provider cooperation.