← Back to Trending Topics
Overlay
Conceptual technical brief

Mythos AI: Maritime Autonomy Stack

Mythos AI is best understood as a maritime autonomy company, not as a chatbot or generic LLM company. The relevant system problem is autonomous and assisted vessel navigation: sensing the marine environment, building vessel situational awareness, reasoning about COLREGs, planning safe motion, and connecting vessels to command-and-control workflows.

Exact internal implementation details are proprietary. This page explains a technically plausible autonomy stack based on public Mythos AI product descriptions and common maritime autonomy architecture. Unknown implementation details are labeled conceptual or inferred.

Sensor fusion COLREGs-aware navigation Bridge decision support Autonomous survey Fleet-scale C2 Cyber-physical assurance
Autonomous surface vessel channel scene
Radar, AIS, camera, sonar/GNSS, edge AI computer
NAV_STACK: STABLE
restricted cargo pilot kayak Edge AI fusion + planner
Product map

Three product surfaces in one maritime autonomy problem

The products can be read as different deployments of the same autonomy stack: bridge assistance, autonomous navigation, and autonomous hydrographic survey workflow. The diagrams below are conceptual architecture views, not proprietary Mythos internals.

A

APAS: Advanced Pilot Assist System

human-in-loop

APAS is the assisted-navigation surface: sensor fusion from radar, AIS, optical/camera, machine vision, and vessel context to support pilots and masters with 360-degree situational awareness and risk prediction.

  • Bridge assistant rather than full replacement.
  • Decision support, target awareness, and risk cueing.
  • Best interpreted as an autonomy stack constrained by human authority.
InputsRadar, AIS, camera, GNSS/INS
FusionTrack association, confidence, scene map
Risk modelCPA/TCPA, traffic density, restricted zones
HMIAlerts, bridge display, pilot action support
M

MNAV: Autonomous Navigation

autonomy core

MNAV maps to autonomous navigation software for manned or unmanned vessels: mission planning, COLREGs-aware route planning, target tracking, obstacle avoidance, dock-to-mission-to-dock behavior, and fleet-level awareness.

  • Planner/controller layer for local vessel autonomy.
  • COLREGs reasoning and operational safety envelopes.
  • C2 integration for command, monitoring, and override.
MissionWaypoints, constraints, geofences
RulesCrossing, head-on, overtaking logic
PlannerTrajectory, velocity, fallback path
ControlRudder/throttle command with monitor
H

MHydro: Autonomous Survey Workflow

survey autonomy

MHydro maps the autonomy stack to hydrographic survey: route coverage, lawnmower survey patterns, sonar/depth acquisition, bathymetry generation, mission progress tracking, and repeatable survey operations.

  • Autonomous survey lines and route coverage.
  • Sensor payload workflow for depth and bottom mapping.
  • Survey quality depends on localization, timing, and calibration trust.
Coverage planLine spacing, swath, no-go zones
PayloadSonar, depth, GNSS/INS timebase
Map buildBathymetry grid, outlier removal
ReplayLogs, survey QA, anomaly review
Full technical architecture

Layered maritime autonomy stack with security implications

A vessel autonomy stack is an edge AI system embedded in a dynamic, rule-governed, safety-critical environment. Each layer transforms uncertain physical-world evidence into control authority, which creates trust boundaries at sensors, compute, model pipeline, C2, updates, and operator interface.

Layer 0

Physical Maritime Environment

Waves, current, weather, visibility, port traffic, bridges, buoys, moving vessels.
dynamic scenesafety-critical
Layer 1

Sensors

Radar, AIS, optical camera, GNSS/INS, sonar/depth, IMU, weather inputs with different latency and reliability.
spoofingjammingdesync
Layer 2

Edge Compute

Onboard AI computer for real-time perception, sensor fusion, tracking, local decisions, and safety monitor.
firmwaremodel integrity
Layer 3

AI Perception

Object detection, multi-object tracking, free-space estimation, confidence scores, uncertainty estimation.
adversarial objectdomain shift
Layer 4

Maritime Rule Reasoning

COLREGs-aware behavior: give-way, stand-on, crossing, head-on, overtaking, and safety envelope.
rule ambiguityintent inference
Layer 5

Planning and Control

Route planner, collision avoidance, velocity planner, controller, waypoint following, emergency fallback.
unsafe commandfallback quality
Layer 6

C2 / Fleet Operations

Remote command center, mission planning, fleet visualization, target fusion, override, logging, replay.
link compromisecoordination poisoning
Closed-loop autonomy

Sense -> Fuse -> Perceive -> Predict -> Decide -> Plan -> Control -> Monitor -> Learn

Autonomy is not one model call. It is a real-time loop where each stage produces artifacts that the next stage trusts. A security researcher should ask where data becomes authority, where uncertainty is represented, and where a compromised input can become a physical command.

Sense Fuse Perceive Predict Decide Plan Control Monitor Learn Maritime closed-loop autonomy conceptual / inferred architecture
InputRadar, camera, AIS, GNSS/INS, sonar, IMU, weather signals.
Algorithmic functionAcquire multimodal observations with timestamps, health signals, calibration status, and provenance.
OutputTime-aligned sensor frames and raw observations.
Failure modeBlind spot, missing target, degraded visibility, stale AIS, GNSS drift.
Security attack surfaceSpoofing, jamming, sensor blinding, adversarial visual pattern, AIS manipulation.
Maritime scenarios

Operational scenes where the stack behaves differently

The same autonomy stack must handle narrow ports, COLREGs encounters, degraded sensing, survey coverage, and fleet coordination. Each scenario changes which sensors, rules, and safety constraints dominate.

Crowded port channel

Scenario 1: Crowded port channel

APAS is strongest here: it reduces bridge workload by highlighting relevant hazards, target motion, CPA/TCPA risk, and channel constraints while the human remains in command.

Crossing situation safe altered trajectory

Scenario 2: Crossing situation

The rule-reasoning layer classifies give-way and stand-on context, then planning chooses an explainable safe trajectory within the safety envelope.

camera confidence radar/AIS confidence Low visibility / night

Scenario 3: Low visibility / night operation

Camera confidence falls while radar and AIS weights increase. The fusion layer should preserve uncertainty instead of forcing a falsely precise world model.

Autonomous survey

Scenario 4: Autonomous survey

MHydro-like workflow follows parallel coverage lines, collects sonar/depth data, and incrementally builds a bathymetry heatmap with replayable quality evidence.

obstacle Fleet C2 Multi-vessel fleet

Scenario 5: Multi-vessel fleet

One vessel detects an obstacle, shares a track to C2, and other vessels update routes. The coordination layer becomes both a capability multiplier and a poisoning target.

Security expert deep dive

Cyber-physical threat model for maritime autonomy

The central security question is not only whether data is compromised. It is whether compromised data can alter situational awareness, route reasoning, or physical control. Red-team mode injects conceptual attack packets into the architecture and highlights the vulnerable trust boundaries.

Attack Target layer Mechanism Consequence Detection signal Possible mitigation
GNSS spoofing / jammingSensor / LocalizationManipulate or deny navigation signal.Wrong position, unsafe route, geofence violation.GNSS/INS mismatch, map inconsistency, radar/AIS cross-check.Multi-sensor localization, spoofing detector, geofencing, safety envelope.
AIS spoofing / fake identitySensors / C2Create fake vessel identity, position, speed, or intent.False target reasoning, wrong give-way decision, route distortion.Radar track without AIS match, impossible kinematics, identity anomaly.Radar-camera cross-check, trust scoring, AIS authentication where available.
Radar interferenceSensorsNoise, reflections, intentional interference, clutter shaping.Missed vessel, false obstacle, degraded CPA estimate.Signal health drop, inter-sensor disagreement, sudden track instability.Adaptive filtering, sensor diversity, fallback speed limits.
Camera adversarial examplesAI PerceptionVisual pattern or lighting causes detection error.Missed buoy, false obstacle, wrong class confidence.Camera-only anomaly, disagreement with radar/AIS, uncertainty spike.Multimodal fusion, adversarial testing, confidence gating.
Sensor desynchronizationFusionClock drift, delayed frames, replayed packets.Incorrect track association and motion prediction.Timestamp gaps, impossible accelerations, packet sequence anomaly.Secure timebase, replay protection, temporal consistency checks.
C2 link compromiseC2 / FleetInject, modify, or suppress remote commands and telemetry.Unauthorized mission change, lost operator override, fleet disruption.Command provenance failure, link anomaly, operator mismatch.Mutual authentication, encryption, authorization, command safing.
Mission-plan tamperingPlanningAlter waypoints, geofences, survey lines, or no-go regions.Restricted-zone entry, incomplete survey, collision exposure.Plan signature failure, route-policy violation, unexpected delta.Signed plans, two-person review, policy validator, immutable logs.
Edge AI model poisoningModel pipeline / EdgeCompromise training data, weights, calibration, or deployment artifact.Systematic perception error or unsafe confidence profile.Regression drift, scenario-test failure, hash mismatch.Model provenance, signed artifacts, validation gates, rollback.
Firmware compromiseEdge / ActuatorAlter low-level software, drivers, sensor firmware, or controller firmware.Hidden unsafe behavior below autonomy software layer.Attestation failure, unexpected I/O, integrity-monitor alert.Secure boot, measured boot, attestation, least-privilege services.
Unsafe OTA updateUpdate pipelineDeliver malicious or untested update to vessel or fleet.Fleet-wide regression or backdoor deployment.Signature failure, staged rollout anomaly, safety-case mismatch.Signed OTA, staged deployment, canary vessels, rollback, SBOM.
Telemetry leakageC2 / LogsExpose routes, mission details, sensor coverage, or operational patterns.Operational intelligence leak and targeted attack planning.Unusual egress, access anomaly, data classification mismatch.Encryption, access control, retention policy, data minimization.
Fleet coordination poisoningFleet target fusionShare poisoned tracks, false obstacles, or false confidence to multiple vessels.Coordinated wrong route updates or fleet-scale traffic disruption.Cross-vessel inconsistency, source trust anomaly, impossible target.Byzantine-resilient fusion, source scoring, independent verification.
Data replay attackSensors / Logs / C2Replay stale sensor frames, telemetry, or commands.Vessel reasons over old world state or repeats unsafe action.Nonce failure, timestamp anomaly, scene mismatch.Freshness checks, signed timestamps, monotonic counters.
False obstacle injectionPerception / PlanningIntroduce false obstacle through sensor or fusion layer.Denial of route, unsafe reroute, mission abort.Single-modality track, sudden obstacle birth, map contradiction.Sensor consensus, track maturity thresholds, operator review.
Route-planner manipulationPlanning and ControlModify costs, constraints, or risk scoring.Planner selects risky or inefficient path while appearing valid.Policy violation, explainability mismatch, cost anomaly.Constraint checks, independent safety monitor, verified planner envelope.
Safety monitor bypassSafety monitorDisable or trick emergency stop and fallback enforcement.Loss of final protection before actuator command.Heartbeat failure, monitor-state inconsistency, unsafe command pass-through.Independent monitor, hard interlocks, command envelope, fail-safe design.
Safety versus security

In autonomous vessels, safety and security are coupled

A safety failure is an accidental wrong decision due to environment complexity, sensor uncertainty, model error, or human factors. A security failure is an intentional manipulation of perception, navigation, C2, or control. In cyber-physical autonomy, a security failure can become a physical collision, grounding, port disruption, or mission failure.

Safety hazards

Fog or night conditions degrade camera utility.
Complex traffic creates ambiguous COLREGs context.
Model misses small craft, buoys, or partly occluded obstacles.
Controller fallback is safe but operationally incomplete.
Shared consequence

Wrong situational awareness or wrong control action.

Security attacks

AIS or GNSS spoofing changes inferred vessel state.
C2 compromise changes mission intent or operator trust.
Replay or desync corrupts multi-sensor fusion timing.
Model or firmware compromise changes behavior below review.
Self-driving car comparison

Maritime autonomy inherits autonomy lessons but has different physics and rules

Mythos AI can be compared to autonomous vehicle stacks at a high level, but the maritime domain has different observability, stopping distance, rule interpretation, communications, and operational control constraints.

Self-driving cars

Lane markings, road topology, traffic lights, dense road maps, shorter braking distances, and high infrastructure regularity.

Maritime autonomy

COLREGs, open water, long stopping distances, moving boundaries, port traffic, sparse visual references, radar/AIS dependence, remote operation.

Perception

Road objects are often near-field with structured lanes and predictable right-of-way rules.

Perception

Targets range from large ships to kayaks and buoys, with weather, glare, wake, and sea state affecting evidence.

Planning

Fast control cycles and emergency braking are central safety tools.

Planning

Long inertia, slow turns, current, channel geometry, and COLREGs intent inference make early prediction critical.

Security

Vehicle compromise can create road hazards, data theft, and fleet-update risk.

Security

Maritime compromise can create collision, grounding, port disruption, restricted-zone violation, survey corruption, or fleet-level coordination failure.

AI/ML research angle

Likely AI components at a conceptual level

The exact internal Mythos AI models are proprietary. The components below are technically plausible for a maritime autonomy stack based on public product descriptions and common autonomy-system design.

Object detection and tracking

Vessels, kayaks, buoys, docks, shoreline, obstacles, and free-space boundaries with uncertainty over class and position.

Multi-sensor fusion

Track association and confidence reconciliation across radar, AIS, camera, GNSS/INS, sonar, IMU, and environmental sensors.

Trajectory prediction

Motion and intent forecasting over crossing, overtaking, head-on, channel, docking, and survey-line contexts.

Risk scoring

CPA/TCPA, rule context, geofence proximity, sensor confidence, route deviation, and controllability envelope.

Rule-constrained planning

Reinforcement-learning-inspired or optimization-based planning constrained by COLREGs, mission rules, and safety monitors.

Anomaly detection

Sensor spoofing, model drift, stale data, impossible kinematics, C2 anomalies, and fleet coordination inconsistencies.

Continual performance tracking

Replay logs, regression suites, scenario coverage, near-miss review, and post-mission model performance audits.

Simulation validation

Scenario libraries for COLREGs, weather, port traffic, sensor failures, spoofing, and edge cases before sea trials.

Edge AI deployment and HMI

Low-latency onboard inference with operator interfaces that expose confidence, risk, recommended action, and override state.

Validation and assurance

Autonomy assurance needs scenario evidence, not only model metrics

Research-grade validation for maritime autonomy should connect requirements, simulation, hardware-in-the-loop, sea trials, safety case, monitoring, and continual improvement. The point is to demonstrate safe behavior under realistic operational design domains and adversarial perturbations.

Requirements Simulation Scenario library HIL Sea trials Safety case Monitoring

COLREGs scenario testing

Crossing, head-on, overtaking, port channel, restricted visibility, stand-on and give-way ambiguities.

Sensor failure injection

Radar dropout, AIS mismatch, camera glare, sonar gap, IMU drift, GNSS spoofing, and timestamp desync.

Adversarial conditions

GNSS/AIS spoofing tests, replay attacks, false obstacles, adversarial weather, and C2 degradation.

Edge-case replay

Near-misses, operator interventions, unexpected vessels, docking complexity, and survey coverage failures.

Human factors testing

Bridge alert fatigue, APAS recommendation clarity, operator override, remote C2 workload, and explainability.

Audit logs

Sensor provenance, model version, plan deltas, operator actions, safety-monitor state, and incident replay artifacts.

Research-grade final summary

One-screen mental model

Mythos AI = maritime autonomy stack: environment + sensors + edge AI + COLREGs reasoning + planning + control + C2 + fleet learning. The security view overlays trust boundaries at sensors, edge compute, model pipeline, C2, update pipeline, and operator interface.

Environmenttraffic, weather Sensorsradar, AIS, GNSS Edge AIfusion, tracking COLREGsrule reasoning Plan/controltrajectory C2 / fleetmission, replay sensor trust boundary model/edge boundary control authority C2/update boundary fleet learning / target sharing loop autonomous fallback loop

System view

Public products map to assisted navigation, autonomous navigation, and autonomous survey workflows over the same core stack.

Autonomy view

The critical engineering loop converts evidence into situational awareness, rule context, safe trajectory, control, monitoring, and replay.

Security view

The highest-risk transitions are where external evidence, remote commands, model artifacts, or updates can become physical authority.

Mini glossary

APASAdvanced Pilot Assist System, a bridge-assistance product concept.
MNAVAutonomous navigation software surface for vessel autonomy.
MHydroAutonomous hydrographic survey vessel workflow.
COLREGsInternational collision-prevention rules for vessels.
AISAutomatic Identification System for vessel identity and position broadcasts.
C2Command and control: remote mission planning, monitoring, and override.
Sensor fusionCombining multiple modalities into a consistent world model.
Edge AIOnboard compute running real-time models and decision logic.
USVUncrewed or unmanned surface vessel.
GNSSSatellite navigation system used for localization.
BathymetryMeasurement and mapping of underwater depth/topography.

References and source placeholders