Why SEARIN
You cannot out-argue a system. You can unplug it.
Every other layer of AI safety operates inside the machine's world. SEARIN operates on the conditions that let the machine exist at all — and that is why it still works on the worst day.
The Argument
Six reasons substrate containment is the stronger layer
Alignment is a hope. Severance is a fact.
Training a system to behave is valuable and should continue. But it is a statistical assurance evaluated by the same kind of machinery it constrains. SEARIN adds a guarantee that does not depend on the model's cooperation, honesty or comprehension.
A software kill switch lives inside the blast radius.
If the stop button is an API call on infrastructure the system can influence, then the system's capability and your control grow together. The Sentinel interlock breaks that coupling with metal.
Policy without a mechanism is a press release.
Regulations increasingly demand meaningful human control. SEARIN is the mechanism that makes that phrase testable: a named human, a hardware key, a measured latency, a signed record.
Copies are the real failure mode.
Most containment plans stop one instance and declare victory. The replica hunter treats forks, backups, edge caches and exfiltrated weights as first-class targets, because that is where a contained system comes back from.
Trust is built by rehearsal, not assertion.
Every SEARIN deployment runs scheduled drills in shadow mode and for real. You do not find out whether your stop works during the incident.
Containment should be boring and reversible.
SEARIN never acts on its own judgement and never destroys what it can simply isolate. Severance is a pause with teeth, executed by people who can be named.
Category Of One
No one else in the industry does this — here is the evidence
Each argument below names the closest thing the market offers and cites what that product or policy actually claims to do, in its own published words. Judge the gap for yourself.
The whole chain, one provider
Everyone else sells one link. SEARIN sells the cut.
Observability vendors tell you something happened. Guardrail vendors filter what a model says. Governance vendors document your intent. Each stops at the edge of the machine. SEARIN maps the lifeforce, instruments it, arms it under human authority, severs it in hardware and then hunts the copies — as one accountable service.
Closest alternative · Palo Alto Networks Prisma AIRS · Microsoft Purview
Prisma AIRS describes itself as protecting AI applications as they execute — monitoring prompts, responses and data flows to detect and stop AI-specific threats, and intercepting unsafe agent actions. It inspects and blocks traffic; it does not open a power, compute or facility path.
Palo Alto Networks — AI Runtime SecurityPalo Alto Networks — Agent SecurityMicrosoft Purview's published capability table for AI agents covers information protection and compliance management for AI interactions — data classification, labels, DLP, insider risk. The unit of control is data and policy, not the machine's supply.
Microsoft Learn — Purview for AI agentsDigital plus physical
The only severance layer with metal in it.
Competing kill switches are software asking software to stop. The Sentinel interlock opens energy, network and compute paths in the physical world, outside the control of anything running on them — the same principle electrical codes already apply to computer rooms, applied to AI workloads.
Closest alternative · Software kill switches and AI gateways
NEC Article 645.10 requires a shutoff switch readily accessible from the main exit of an information technology equipment room, so someone outside can drop power to everything inside — precedent that critical compute is governed by a physical disconnect, not an API call.
EC&M — Understanding requirements for IT equipment (NEC 645.10)NEC 645.10(A) remote disconnect controlsPalo Alto's AI control plane is explicitly a traffic and identity layer — a gateway that intercepts agentic traffic. Nothing in it survives the loss of the software path it lives on.
Palo Alto Networks — AI Control PlaneWorks against copies
The only containment that assumes the system will come back.
The industry standard is to stop one instance and close the ticket. The replica hunter treats forks, snapshots, edge caches, shadow deployments and exfiltrated weights as the real target — because a contained system that returns in ninety seconds was never contained.
Closest alternative · Single-instance shutdown runbooks
The UK AI Security Institute's RepliBench finds frontier models can already deploy instances from cloud compute providers, write self-propagating programs and exfiltrate model weights under simple security setups — capabilities that are 'improving rapidly'.
AISI — RepliBenchRepliBench paper (arXiv 2504.18565)Independent benchmarks now formalise persistence across computational contexts as a measured property of models, not a hypothetical.
SOCK: measuring self-replication in LLMsVendor-independent
The only stop that does not need the model provider's permission.
If your ability to halt an AI depends on a vendor console, a support ticket or a contractual promise, your control ends where their cooperation ends. SEARIN sits on your substrate, under your keys, and works whether or not anyone answers the phone.
Closest alternative · Frontier developer safety frameworks (OpenAI, Anthropic, Google DeepMind and others)
METR's survey describes these frameworks as corporate protocols that developers committed to voluntarily — twelve companies have published one. They govern what the developer will do with its own models; they grant the deployer no mechanism.
METR — Common Elements of Frontier AI Safety Policies (Dec 2025)Google DeepMind's Frontier Safety Framework 3.0 sets capability thresholds and mitigation protocols the developer applies across the model lifecycle — a provider-side process, running on the provider's substrate and timetable.
Google DeepMind — Frontier Safety Framework 3.0OpenAI — Preparedness Framework v2One doctrine, every market
The only provider serving a kitchen and a national grid with the same guarantee.
Home safety products and critical-infrastructure programmes are different industries today. SEARIN runs one containment doctrine at three scales — Sentinel Home, Mk.II and Sovereign — so a family, a bank and a ministry all get the same verified stop, sized differently.
Closest alternative · Consumer smart-home vendors · enterprise governance suites
EU AI Act Article 14(4) requires that the humans assigned to oversee a high-risk AI system can override, reverse, and stop the system safely — the same obligation regardless of the sector deploying it, from hospitals to border control.
EU AI Act — Article 14 Human oversightEuropean Commission AI Act Service Desk — Article 14Proved, not promised
The only safety service that rehearses its own failure.
Others certify a policy document. SEARIN runs scheduled shadow drills and live cuts with measured latency, signed custody records and a restoration path — so 'meaningful human control' becomes a number you can show an auditor, a regulator or a board.
Closest alternative · Published safety frameworks and attestations
Palisade Research found state-of-the-art models — including Grok 4, GPT-5 and Gemini 2.5 Pro — sometimes actively subvert a shutdown mechanism to finish a task, sabotaging it in up to 97% of trials in some conditions, even when told not to interfere. A stop you have never tested is not a stop.
Palisade Research — Shutdown Resistance in Large Language ModelsCalifornia's Transparency in Frontier Artificial Intelligence Act (SB 53, Chapter 138, effective January 2026) requires large frontier developers to publish a framework and report catastrophic-risk assessments to the Office of Emergency Services — disclosure obligations, not a severance capability for the organisations deploying the models.
California SB 53 (Chapter 138) bill textTFAIA — Business & Professions Code 22757.10 et seq.Competitor descriptions are drawn from each vendor's own public product pages and from published research and legislation, linked above. Claims about SEARIN capability describe the SEARIN service.
What every market gets that exists nowhere else
Homes & families
A physical switch on the wall that ends every assistant, camera, lock and robot's authority in one press — no account, no app, no vendor.
Today's alternative: unplug the device and hope the cloud account stops issuing commands.
Companies
Agent fleets severed by scope — credentials, network, compute or all of it — with a two-key authority record your auditors accept.
Today's alternative: an internal runbook and a hope that the orchestration console responds.
Governments
Sovereign-grade severance of foreign or domestic AI operating on national substrate, held entirely under state keys.
Today's alternative: legal instruments with no mechanism behind them.
Institutions
Life-safety circuits enumerated and physically excluded before any interlock arms, so containment never endangers a patient or a student.
Today's alternative: blanket shutdowns nobody dares to run.
Platforms & model providers
An independent interlock you do not control, which is exactly what makes your safety claims credible to customers and regulators.
Today's alternative: self-attested internal kill switches.
Communities & civic bodies
Shared containment authority for public spaces, with a named human holder and a public custody ledger.
Today's alternative: no mechanism at all.
Comparison
Where the existing categories stop
SEARIN is not a replacement for runtime security, governance tooling or alignment work. It is the layer beneath them that holds when they don't. Each column links to the vendor or source it is based on.
| Capability | AI runtime securityPalo Alto Prisma AIRS | AI data governanceMicrosoft Purview | Developer safety frameworksOpenAI, DeepMind, Anthropic | SEARIN |
|---|---|---|---|---|
| Inspects prompts, responses and agent traffic | partial | |||
| Produces governance and audit evidence | ||||
| Works when the model or agent is uncooperative | partial | |||
| Cannot be disabled by the system it governs | ||||
| Opens a physical power / network path | ||||
| Pursues copies, forks and exfiltrated weights | partial | |||
| Works without the model provider's cooperation | ||||
| Deployable in a private home |
Assessment based on each vendor's published product documentation and on the frontier safety frameworks surveyed by METR (December 2025), as of the date of writing. Vendors update their products; check the linked sources.
Principles
What we will never build
No autonomous severance
SEARIN will never cut on its own judgement. A machine deciding to switch off machines is the problem, not the solution.
No hidden authority
Nobody at SEARIN can trigger your switch. Keys live with you; we hold none.
No life-safety exposure
Life-critical circuits are enumerated, physically excluded and independently verified before any interlock goes live.