AI CONTROL · Pitch deck · AI Future track, Blockchain Life 2026
1 / 12←→Download PDF
AI CONTROL

ControlCore · AI Future track · Blockchain Life 2026 · Dubai · 1–2 December

Human authority over autonomous intelligence.

AI may be autonomous in operation, but autonomy does not create authority. ControlCore puts a human-authorized decision in front of every consequential action.

aicontrol.si1 / 12

The problem

Autonomy arrived. Authority never did.

Agents, copilots and robot planners now act on production systems, funds and the physical world. Identity systems answer who is asking. Nothing answers whether it may.

Failure modes the platform is built around

credential theftcompromised developerprompt injectionprivilege escalationunknown AI channelunknown machine languagecompromised robotinsider misusepolicy misconfigurationtelemetry manipulationcloud outagetrust service outage
AI CONTROL · The problem2 / 12

The thesis

Authority is granted. Never assumed.

No authority by default

Unknown agents are blocked. An agent nobody registered is an agent nobody authorized.

Identity is not authorization

Knowing who an agent is says nothing about what it may do. Capability is not permission either.

Revoked stays revoked

A revoked or quarantined agent is blocked everywhere, at once, including at the edge.

High impact needs a human

Production writes, robot motion, physical actions and funds always wait for a person. Policy only ever makes a decision stricter.

AI CONTROL · The thesis3 / 12

ControlCore

Every request runs the same gauntlet

Each stage can only narrow what the previous one allowed. Every verdict is signed and kept in an append-only audit record.

REQUESTAn agent, copilot or robot planner asks to act.
IDENTITYWho is asking, proven, not a name in a header.
CAPABILITYCan this agent perform this class of action at all?
POLICYVersioned, signed, auditable rules.
CONTEXT & RISKTarget, blast radius and live telemetry become a score.
HUMAN APPROVALHigh impact pauses for an authorized person.
VERDICTSigned, then executed exactly as approved.
ALLOWProceed through the execution gateway.
LIMITProceed with a narrower scope or rate.
QUARANTINEIsolate the agent, keep the evidence.
BLOCKDeny, and raise it to the SOC.
AI CONTROL · ControlCore pipeline4 / 12

The platform

One control plane, many enforcement points

ControlCoreThe pipeline, as one authorize endpoint.
AI FirewallInspect and enforce every AI request in flight.
Policy EngineVersioned, signed, auditable policy.
Risk EngineContextual risk as an input to policy.
AI PassportEvery agent's identity and authority, committed.
Trust ChainVerifiable authority, anchored on Solieum.
Communication IntelligenceAI-to-AI traffic, decoded for humans.
Robot GatewayAuthorized commands only reach physical systems.
EdgeLocal enforcement that survives outages.
SOCSecurity operations for autonomous systems.
AI CONTROL · Platform5 / 12

Communication Intelligence

When humans cannot read the machine language

Agents talk to agents. The traffic is visible to authorized humans, and the AI Language Decoder turns it into something a person can judge.

Decode

Protocol discovery and machine-language parsing turn signal into structure.

Explain

Semantic intelligence turns structure into meaning, and behavior correlation shows what happened next.

Verify

A communication graph with replay, and forensic evidence records with integrity guarantees.

AI CONTROL · Communication Intelligence6 / 12

Trust Chain on Solieum

Prove every decision.

Identities, policies and control decisions are anchored on Solieum, the Layer 2 for Solana. Anyone holding a proof ID can check who authorized what, and when, without trusting the operator's word.

RegistriesAgents, policies and approvers, committed.
AnchoringDecision records rolled up and settled on Solana through Solieum.
VerificationPaste a proof ID into the Proof Explorer and check it yourself.

Why a chain at all: an audit log the operator controls can be rewritten by the operator. An anchored one cannot.

AI CONTROL · Trust Chain7 / 12

Operating modes

From human-only to controlled autonomy

Autonomy is earned per agent, per system, through policy. The high-impact classes keep a human in every mode.

Human-onlyThe agent proposes. A person performs every action.
Human-approvedThe agent acts after an authorized person approves each request.
SupervisedLow-impact actions proceed under policy. Everything else waits.
Controlled autonomyPolicy, risk and limits decide. Humans still own production, funds and motion.
AI CONTROL · Operating modes8 / 12

Where it applies

Wherever AI touches something that matters

Enterprise AI & copilots Autonomous agents Robotics Humanoid systems Industrial systems & OT Cloud infrastructure Financial & digital asset infrastructure Data centers Healthcare Government & critical infrastructure AI research labs Multi-organization AI governance

Exchanges, trading platforms and market infrastructure are the digital-asset case: AI in payments, treasury and trading operations, with funds movements behind a human.

AI CONTROL · Where it applies9 / 12

Developer platform

Built to be integrated, not bolted on

REST API

One authorize endpoint under /api/v1, with an OpenAPI 3.1 specification and signed webhooks for event delivery.

SDKs

TypeScript and Python today. Rust and Go planned.

Runs anywhere

Node backend, PostgreSQL and Redis, Docker Compose for the full stack, and an edge mode that keeps enforcing through an outage.

AI CONTROL · Developer platform10 / 12

Where it stands

Labelled honestly, stage by stage

AI Control Core · stage: Development
Concept

Described in the white paper.

Research

Being validated.

Development

Being built and tested.

Testnet-Pilot

Running with pilot partners.

Planned

On the roadmap.

Every feature on aicontrol.si carries one of these labels. The white paper is the canonical source. Nothing is offered as production infrastructure yet.

AI CONTROL · Status11 / 12

AI CONTROL

Intelligence should not mean unlimited authority.

Scan for the AI CONTROL brochure
solieum.com/brochures/ai-control

© 2026 AI CONTROL · Not production infrastructure12 / 12