1 of 7

The AI Trust layer for the US Navy.

BUILDING AN AI-FIRST MARITIME FORCE

OpenAI

Grok

Claude

Gemini

GenAI.mil

The question is no longer whether the Navy will use AI. It is what happens when that AI loses the network.

Autonomous Fleets

2 of 7

At sea, connectivity is a variable not a dependency.

STRUCTURAL

mast/superstructure blockage

EMCON

intentional emissions silence

ELECTRONIC WARFARE

jamming/contested spectrum

ENVIRONMENT

weather/propagation/ orbital geometry

NETWORK

congestion/outage/ degraded bandwidth

LIVE

DEGRADED

DENIED

A cloud-dependent agent fails exactly when the environment becomes most contested.

3 of 7

What happens when we lose connectivity?

CLOUD-DEPENDENT AGENT

CLOUD

POLICY / IDENTITY

EXECUTION

DEADLIGHT

LOCAL

POLICY

DEADLIGHT

GATE

CANDIDATE

STATE

TOOL

REQUEST

CLOUD

POLICY / IDENTITY

Deadlight makes connectivity optional — not safety.

UPLINK LOST

AGENT

TOOL

REQUEST

NETWORK

AGENT

SANDBOX

AGENT

NETWORK

EXECUTION

ALLOW

DENY

Policy unavailable

Identity unavailable

Tool authorization unavailable

Agent behavior undefined

4 of 7

Deadlight: The trust layer between AI and the real world.

MODEL-AGNOSTIC

Supports OpenAI, Grok, Claude, local models

SANDBOXED

Isolated yet high fidelity agent execution environments

POLICY-AWARE

Locally cached, signed, versioned policies

FAIL-CLOSED

Unsafe actions blocked before system impact

FRONTIER MODEL

DEADLIGHT

The trust layer

REAL WORLD

5 of 7

We don't trust the model.

We restrain the execution.

FRONTIER MODELS

Local LLMs

tool / action proposals

MODEL AGNOSTIC

frontier and local models submit tool and action proposals.

DEADLIGHT HARNESS

agent

runtime

tool

registry

context

manager

policy

engine

EXECUTION CONTROL

orchestrates agents, tools, context, and policies under strict rules.

LOCAL VECTOR MEMORY

embeddings for mission context

signed

policies

telemetry

VERIFIED CONTEXT

vector embeddings ground mission knowledge and signed policies.

DAYTONA SANDBOX

isolated

filesystem

controlled

processes

tool & network

isolation

ephemeral

SAFETY GATE

ALLOW

DENY

OpenAI

policy

validation

candidate-state

simulation

anomaly

detection

risk

evaluation

capabilities

Grok

Claude

ISOLATED ENVIRONMENT

commands run in an ephemeral sandbox with no persistent side effects.

DECISION ASSURANCE

multi-layer checks validate actions and evaluate risk before anything executes.

6 of 7

LIVE DEMO: VIDEO

https://youtu.be/AHLEeDa9k-k

7 of 7

The AI-first fleet needs an offline trust layer.

OpenAI

Local Models

DEADLIGHT

Trust Layer

Ship / Edge Compute

Operational Data

Autonomy Agents

Logistics Agents

Maintenance Agents

Intelligence Agents

CX

Agents

MODEL-AGNOSTIC

Compatible with Navy-authorized models

PLATFORM-AGNOSTIC

Works alongside existing ship/edge software

MISSION-AGNOSTIC

Protects diverse agent types

Designed to complement emerging Navy edge-data and autonomy infrastructure.

OFFLINE. SANDBOXED. POLICY-BOUND.

Keep the intelligence up, always.

Grok

Claude