Utility Cybersecurity and Trusted Industrial AI | UtilityGrid
UtilityGrid · Research

Cybersecurity, Physical Security, and Trusted Industrial AI for Utilities

Connected utility systems depend on trustworthy observations, controlled access, and authorized actions. As artificial intelligence (AI), the Internet of Things (IoT), and Physical AI become part of utility operations, cybersecurity must extend beyond enterprise networks to the devices, sensors, control interfaces, mobile platforms, and physical assets that support essential services.

Explore Research ↘
⌁
Trust and Security Across Critical Utility Infrastructure Identity · Sensors · Networks · Actions
Operating Context

Why Utility Security Requires an Operational View

UtilityGrid examines security and resilience questions across electric power, gas, water, and wastewater environments. Relevant operating settings include substations, distribution networks, generation and storage facilities, pumping stations, treatment plants, metering infrastructure, and remote field assets.

Each environment has different safety obligations, equipment lifecycles, connectivity limits, and consequences when systems are unavailable.

A security control that is appropriate for an office network may not be suitable for a protection relay, remote terminal unit (RTU), programmable logic controller (PLC), smart meter, pump controller, or substation gateway.

The research focus is the complete path from physical observation to operational decision and verified outcome. That path includes device identity, sensor integrity, network boundaries, AI model and data integrity, command authorization, physical access, privacy, incident recovery, and human oversight.

The objective is not to connect every device or automate every response. It is to make each relevant data flow and action path understandable, limited to its intended purpose, and testable against operational requirements.

Research Priorities

Building Trust Across Utility Systems

Research examines how utility infrastructure can establish trustworthy device identity, protect operational networks, preserve sensor integrity, and govern AI-supported decisions.

01

Device Identity, Secure Onboarding, and Lifecycle Management

Utility networks include field sensors, intelligent electronic devices, gateways, meters, cameras, controllers, and mobile inspection equipment. Research examines how each device is associated with the correct physical asset, authenticated before receiving access, and assigned appropriate permissions when replaced, moved, serviced, or retired.

Candidate methods include hardware-backed credentials where supported, certificate and key lifecycle management, device attestation, secure boot, controlled onboarding, credential revocation, and auditable ownership changes.

A readable RFID tag or device label can help identify an asset, but it does not by itself prove that a device is authentic. Research also considers replacement sensors and temporarily disconnected remote devices.

02

OT Network Segmentation and Restricted Remote Access

Substations, generation facilities, water treatment plants, pumping stations, and distribution control environments may contain equipment with different communication and maintenance requirements.

Research considers how to separate information technology (IT), operational technology (OT), safety-related, and third-party service paths while preserving approved diagnostics and maintenance workflows.

Candidate methods include segmented network zones, authenticated gateways, narrowly scoped remote sessions, time-limited credentials, role-based permissions, and auditable access approvals.

Evaluation should consider unauthorized access paths, policy errors, revocation time, latency, and the effect of controls on legitimate maintenance and operating tasks.

03

Sensor Integrity and Trustworthy Utility Telemetry

AI-supported monitoring is only as reliable as the observations it receives. Voltage, current, temperature, pressure, flow, vibration, water quality, equipment status, location, and time data can be affected by sensor faults, poor calibration, communication delays, or malicious manipulation.

Research examines message authenticity where supported, timestamp and sequence checks, sensor-to-asset binding, independent measurements, and physical consistency checks.

An integrity alert should not automatically be treated as proof of malicious activity. Ordinary equipment faults and operating changes can produce similar symptoms.

04

Trusted Industrial AI and Command Authorization

An AI assistant or agent may summarize alarms, interpret operating records, recommend an inspection, or prepare a command request. These functions do not grant authority to operate equipment.

Candidate controls include external policy enforcement, asset-specific permissions, command allowlists, approval records, freshness checks, parameter validation, scoped credentials, and independent authorization at the control boundary.

Approval should be bound to the intended asset, action, parameters, and time window. The system must distinguish a recommendation, a work-order request, a human-approved command, and a command that has actually been executed.

Security Architecture

Four Layers of Trusted Utility Operations

A secure utility workflow must consider the complete relationship between physical assets, operational observations, digital intelligence, and authorized actions.

01

Device Identity

Establishes the identity of connected devices and associates them with the correct physical assets, credentials, ownership, and lifecycle records.

02

Trusted Observations

Examines measurement authenticity, timestamps, sequence integrity, sensor quality, asset associations, and independent physical consistency.

03

AI and Decision Integrity

Protects model versions, data provenance, recommendations, approval records, permissions, and decision boundaries.

04

Authorized Physical Action

Requires defined permissions, independent authorization, controlled command dispatch, and verification of the resulting equipment state.

Core principle: Device identity does not prove measurement accuracy. A plausible measurement does not prove its source is trustworthy. An AI recommendation does not establish operational authority.
⌁

Utility Substation Physical Security and Asset Protection

Access monitoring · Perimeter protection · Asset custody · Human review

Physical Infrastructure

Physical Access, Perimeter, and Asset Security

Utility physical security includes controlled access to substations, switchyards, control rooms, pumping stations, treatment facilities, storage areas, and remote cabinets.

Research can examine how access events, credential status, camera observations, tamper sensors, and asset movement records are correlated to support security review.

Candidate methods include identity-aware access records, monitored entry points, tamper-event correlation, asset custody checks, and human-reviewed anomaly triage.

Busy shift changes, contractor access, emergency entry, and authorized maintenance can create patterns that resemble exceptions. A single camera event or proximity observation should not be treated as conclusive evidence of misconduct.

Any access-control response must preserve emergency egress, worker safety, and site operating procedures. Physical security analytics should support accountable review rather than automatic conclusions about an individual.

Trusted Industrial AI

From AI Recommendation to Authorized Command

AI-generated recommendations must remain separate from the authority to operate physical equipment.

Intelligence must never replace operational authorization.
01

Recommendation and Work-Order Request

An AI assistant may summarize alarms, interpret operating records, identify an inspection requirement, or prepare a command request. These outputs remain proposals until reviewed and authorized through the appropriate process.

02

Independent Policy Enforcement

Candidate controls include external policy enforcement, asset-specific permissions, command allowlists, scoped credentials, parameter validation, and approval records.

03

Asset and Action Validation

Approval should be bound to the intended asset, action, parameters, and time window. A change in equipment state or permissions may require the request to be reviewed again.

04

Execution and Physical Verification

A command acknowledgment from an interface is not proof that physical equipment reached the intended state. Independent feedback and human oversight remain important.

Privacy and Data Protection

Protecting People, Data, and Operational Records

Utility monitoring systems can collect sensitive operational and personal information. Security research must consider the usefulness of a workflow alongside the information it collects, retains, and exposes.

01

Data Minimization

Collect only the information required for a documented operational purpose. Research examines purpose-specific access, pseudonymous identifiers, edge processing, and appropriate retention limits.

02

Access and Auditability

Access to sensitive video, location records, access logs, and workforce information should be controlled and auditable. Records should capture appropriate viewing and export activities.

03

Workforce Privacy

Workforce monitoring should be limited to a documented operational purpose and reviewed against applicable privacy and labor requirements.

Privacy evaluation: Pseudonymization and aggregation can reduce exposure, but they should not be described as eliminating reidentification risk.
Resilience and Recovery

Incident Containment, Recovery, and Verified Restart

Cyber incident recovery requires more than restoring a server or reconnecting a gateway. Teams must establish whether device identity, software versions, sensor reliability, equipment position, and queued commands can be trusted.

01

Incident Containment

Research examines how teams can contain a suspected incident while preserving essential utility functions and maintaining an accurate incident record.

02

Trusted Configuration Restoration

Candidate methods include trusted configuration backups, credential replacement, staged restoration, and rollback to known configurations.

03

Physical State Reconciliation

Teams may need to reconstruct the current physical state, reconcile equipment records, review queued commands, and independently check equipment condition.

04

Verified Restart

A restart decision should account for process consequences and the equipment’s actual state. Disconnecting a system or restarting it automatically may not be appropriate for every operating environment.

Mobile and Autonomous Systems

Mobile Platforms, Secure Teleoperation, and Fleet Integrity

Some utilities use or evaluate drones, remotely operated inspection equipment, and mobile robots for asset inspection or site monitoring.

01

Secure Teleoperation

Research considers authenticated command links, secure updates, current position and mission state, safe handover between local and remote control, and the integrity of shared maps or fleet status.

02

Fleet Integrity and Safe Operation

A mobile platform should not accept obsolete or unverified commands after a communications interruption. A compromised fleet member should not make other units rely on unverified position, mission, or map information.

Testing boundary: Tests should use simulators or isolated, authorized platforms and should avoid interference with public navigation or communications services.
Software and Cryptographic Security

Model, Data, and Software Supply-Chain Integrity

Industrial AI depends on more than a model file. Training data, labels, preprocessing, dependencies, firmware, configuration, model versions, and deployment approvals all affect system behavior.

01

Model and Data Provenance

Research examines provenance records, artifact verification, controlled updates, dependency inventories, evaluation before release, and rollback to a known configuration.

A useful audit trail should connect a model output to relevant data sources, model version, configuration, decision record, approval, and any subsequent command.

02

Cryptographic Agility and Post-Quantum Migration

Long-lived utility equipment creates a planning challenge as cryptographic requirements evolve. Research can assess how gateways and constrained field devices support algorithm updates, credential replacement, signed software delivery, and interoperability across mixed equipment generations.

Post-quantum cryptography should be evaluated against the resource limits of representative gateways and embedded devices.

Technical reference: NIST FIPS 203 – Module-Lattice-Based Key-Encapsulation Mechanism (ML-KEM).

NIST FIPS 204 – Module-Lattice-Based Digital Signature Standard (ML-DSA).

Relevant evaluation measures include handshake and verification latency, memory and storage use, message size, energy demand, update recovery, and effects on time-sensitive workloads.
Physical AI Architecture

From AI Decision to Verified Physical Outcome

Security controls support the identification, sensing, decision, and action functions used in Physical AI and Artificial Intelligence of Things (AIoT).

01

Identification Engine

Associates an observation with the correct device or physical asset and its relevant identity context.

02

Sensing Engine

Captures measurements and their quality context, including source integrity, timing, and physical consistency.

03

AI Decision Engine

Evaluates the information and produces a recommendation or decision proposal within defined permissions.

04

Physical AI Action Engine

Connects an approved request to an equipment interface and checks the resulting physical state.

Trust boundary: Each transition requires a defined trust boundary. A recommendation is not authorization. A command acknowledgment is not proof of physical completion. Independent feedback and human oversight remain important when outcomes affect service continuity, worker safety, or equipment condition.
Proposed Evaluation

Proposed Demonstration and Evaluation

A suitable starting point is an isolated, authorized testbed representing a defined utility workflow.

Depending on the selected research question, the testbed may include simulated field devices, a gateway, representative telemetry, a control-system interface, a mobile-platform simulator, or a bounded hardware rig. No test should be conducted against live utility equipment or production networks without explicit authorization and an approved safety plan. A demonstration can compare a documented baseline with a candidate control or workflow. Test conditions may include invalid device identity, stale or conflicting measurements, denied authority, unapproved software changes, interrupted communications, and recovery from a known configuration.

Evaluation Measures

Evaluation measures should be defined before testing and selected according to the workflow being assessed.

↗

Unauthorized-action success rate and the number of invalid requests rejected.

↗

Detection delay, false-alarm rate, and missed integrity or access events.

↗

Legitimate-task completion and any degradation in permitted maintenance or operating work.

↗

Credential revocation and policy-change propagation time.

↗

Recovery time, state reconciliation completeness, and incorrect restart prevention.

↗

Physical outcome checks that independently confirm whether the intended equipment state was reached.

↗

Resource cost of cryptographic changes, including latency, memory, storage, energy, and message size.

Validation and Safety Boundaries

Results should identify the test configuration, baseline, data sources, assumptions, run count, and limitations. Synthetic data must be labeled as synthetic. If no physical reference data is available, describe the work as plausibility-reviewed rather than field-validated. Report adverse outcomes and cases where a proposed control reduces availability or legitimate task performance.

Research Status and Scope

Research Status, Scope, and Evidence

Clear evidence boundaries are essential when describing cybersecurity and trusted industrial AI research for utilities.

!

Proposed Research Direction

This page presents a Utility-focused adaptation of a shared research agenda. The shared initiative is associated with Aperture Physical AI Research Labs and the work of Aperture Venture Studio, GAO RFID Inc., and GAO Tek Inc.

Shared research does not, by itself, establish a separate UtilityGrid research team, a completed UtilityGrid product, a customer deployment, or a measured utility-sector result.

The topics described here are research directions and candidate evaluation methods. UtilityGrid-specific experiments, demonstrations, staffing, and operational results should be identified only after the responsible technical owner confirms the evidence.

Research evidence is not a cybersecurity certification, safety certification, regulatory approval, or proof that a control is suitable for every utility environment. Site-specific engineering, operational approval, and validation remain necessary.

Start a Conversation

Discuss a Utility Security Research Question

Start with one asset class or workflow, such as substation access monitoring, remote sensor identity, control-room decision support, or verified recovery after a communications interruption. UtilityGrid can use that scope to identify relevant data, interfaces, operational constraints, and evaluation criteria.

Explore Research ↗