Cybersecurity and Physical Action Safeguards for Utilities
Explore cybersecurity and execution safeguards for utility systems, including permissions, oversight, network segmentation, and outcome verification.
Explore the safeguards ↘Why utility cybersecurity must account for physical operations
Connected utility systems need controls for both information access and operational execution. A sensor reading, AI recommendation, remote maintenance session, or equipment command can affect how a substation, feeder, pump, or other utility asset is monitored and operated.
The security design must protect devices and data while defining who can request an action, what conditions permit it, how protective functions remain independent, and how the resulting physical state is confirmed.
Security must reflect OT performance, reliability, and safety
Supervisory control and data acquisition (SCADA) systems, remote terminal units (RTUs), programmable logic controllers (PLCs), intelligent electronic devices (IEDs), protection relays, sensors, gateways, and engineering workstations may participate in utility monitoring or control workflows. These systems have availability, timing, reliability, and safety requirements that differ from conventional enterprise IT.
Monitoring and sensing
Connected devices identify assets, collect measurements, and provide operational visibility.
AI-supported decisions
Analytics can support decisions, but input integrity, asset association, and data freshness remain essential.
Authorized execution
Connected command paths require defined authority, equipment boundaries, operator approval, and independent outcome checks.
NIST reference: NIST SP 800-82 Rev. 3 remains a published reference. The September 2026 SP 800-82 Rev. 4 document is an initial public draft and should be treated as draft guidance pending finalization. Read NIST SP 800-82 Rev. 3 or review the September 2026 SP 800-82 Rev. 4 draft.
Map the complete utility OT system boundary
Start with an asset and data-flow inventory showing how field equipment, control networks, enterprise applications, external service providers, and administrative users connect. Include physical and logical paths used to monitor assets and request or execute actions.
- Field and substation devices: sensors, meters, RTUs, PLCs, IEDs, protection relays, actuators, and local control panels.
- Control and operations systems: SCADA servers, HMIs, engineering workstations, historians, and control-center applications.
- Communications infrastructure: field networks, utility WAN links, cellular or radio backhaul, gateways, firewalls, remote-access services, and network management systems.
- Enterprise and analytics: asset management, outage management, work management, data platforms, AI services, and reporting applications.
- External and physical access: vendor support, integrators, cloud services, maintenance laptops, control-room entry, substation perimeter controls, cabinet access, removable media, and maintenance procedures.
Illustrative OT zones and conduits
SCADA · RTUs · PLCs
Sensors · IEDs
Document the purpose, owner, trust level, permitted data flows, and consequence of compromise for each component. Identify read-only systems, workflow-request systems, and equipment-command systems. Treat undocumented connections and shared administrator accounts as items requiring review. Network segmentation should separate enterprise services from control environments, define required conduits, and use a DMZ or controlled intermediary where appropriate. Test segmentation so protection, monitoring, and recovery functions are not disrupted.
Apply identity, least privilege, and protected access
Every user, service, device, and application accessing utility OT should have a known identity and assigned purpose. Use individual accounts where supported, role-based permissions, strong authentication, and time-limited access for maintenance or vendor work.
Control who and what can connect
- Grant the minimum privileges needed for a defined task, then remove or expire access when the task ends.
- Require multifactor authentication for remote and privileged access where technically and operationally feasible.
- Separate operator, engineer, administrator, vendor, and service identities; avoid shared credentials and standing privileges.
- Use approved procedures for provisioning, credential rotation, emergency access, and revocation.
Make remote sessions accountable
- Use managed jump hosts or bastion systems as controlled entry points for approved remote sessions.
- Restrict destinations, permitted tools, and session duration; disable remote access when it is not needed.
- Protect communications with encryption and authenticated endpoints, confirming OT protocol support.
- Record session times, identity, source, target asset, approval, actions, and configuration changes.
Remote access should be enabled only through an approved process, monitored while active, and closed after authorized work is complete. CISA’s industrial remote-access guidance emphasizes defense in depth and tailoring remote-access arrangements to the control-system architecture. See CISA guidance on configuring and managing remote access for industrial control systems.
Protect devices, software, and operational data
Device security begins before installation and continues through retirement. Maintain an inventory of firmware and software versions, ownership, network location, support status, and update responsibility. Verify device identity during onboarding and prevent unknown or unapproved equipment from joining trusted networks.
- Approve firmware, software, configuration, and AI model updates through a controlled process; validate them in a representative test environment.
- Maintain secure configuration baselines for gateways, servers, engineering workstations, and supported field devices.
- Protect configuration files, credentials, model artifacts, event histories, and command records against unauthorized alteration.
- Monitor unexpected device enrollment, configuration changes, unusual communications, failed authentication, and abnormal command activity.
- Back up configurations and essential operational data; test restoration procedures and document recovery dependencies.
- Coordinate patching with vendors and operations teams; document compensating controls and residual risk when immediate patching is not feasible.

For AI-supported workflows, record the model or service version, input data source, timestamp, quality status, decision output, and the person or system approving downstream use. A compromised sensor, stale measurement, incorrect asset association, or altered model artifact can undermine an otherwise valid decision process. Treat input integrity and data provenance as part of the security boundary.
Separate software permissions from equipment protection
Application permissions and command validation do not replace engineered equipment protections or safety functions. Preserve the independence and authority of equipment-level interlocks, protection relays, safety instrumented functions where applicable, and approved control logic.
Define the rules before connecting decisions to equipment
Before connecting an AI-supported recommendation to a physical action, define the target asset, permitted command set, operating envelope, required approvals, and conditions that require rejection or escalation. Supervisory commands should remain distinct from low-level control loops and protection functions.
Validate every action request
- Use an explicit allowlist of commands and target assets rather than arbitrary command text.
- Validate asset identity, command parameters, operating mode, interlocks, user authority, and freshness of observations before dispatch.
- Require human approval for actions whose risk, novelty, or operating context exceeds the approved automation boundary.
- Define rate limits, duplicate-request handling, timeout behavior, retry limits, and safe handling of partial execution.
Preserve operator authority and verify state
- Provide a clear operator override and a documented method to return the workflow to a known operating state.
- Verify completion using independent observations, such as a subsequent device state, sensor reading, position measurement, or operator confirmation.
- Keep supervisory commands separate from low-level control loops and protection functions.
- Do not bypass local controls, protection settings, or established operator procedures.
Acknowledgment is not proof of physical completion. A command acknowledgment confirms that a message was received or accepted by an interface. It does not independently establish that a breaker, switch, valve, or other asset reached the intended state. Record the requested state, authorization decision, dispatched command, acknowledgment, observed outcome, and any exception as linked but distinct events.
Plan for stale data, invalid commands, and communication loss
A resilient design defines what happens when inputs or communications cannot be trusted. It should not assume every sensor reading is current, every command is valid, or every network path remains available.
Stale observations
Reject or quarantine decisions that rely on measurements older than the approved age limit for that workflow. Escalate when freshness cannot be established.
Invalid or out-of-range commands
Block requests that fail asset, authority, parameter, mode, or interlock validation. Preserve the rejection reason in the audit record.
Compromised or inconsistent inputs
Compare readings with independent sources or plausibility rules where available. Flag unexplained disagreement for operator review.
Communications loss
Define whether each function continues locally, buffers noncritical records, degrades to advisory mode, or stops accepting new action requests.
Delayed or duplicate messages
Use unique command identifiers, expiry windows, and idempotent processing where supported to prevent unintended repeat actions.
Partial execution and restart
Detect mismatch between requested and observed state, stop automated retries where appropriate, and verify asset state, communications health, configuration integrity, and outstanding commands before restoring normal command privileges.
Fail-safe behavior is application-specific. A generic instruction to stop all equipment may itself create risk in a utility environment. Define the appropriate degraded mode with control engineers, protection specialists, operations personnel, and safety owners for each asset class and operating scenario.
Implementation checklist for utility project teams
Use this checklist during architecture review, pilot planning, commissioning, and periodic reassessment. Assign an accountable owner and retain evidence for each item.
Validate safeguards before enabling physical action
Begin with an isolated, authorized test environment that represents the intended utility workflow. Use simulated devices, a command gateway, or a bounded hardware testbed to examine denied authority, manipulated observations, stale measurements, invalid parameters, compromised update packages, communications interruption, and recovery.
Test safely
Do not test disruptive actions on live utility equipment without formal authorization, engineering review, and an approved test plan.
Measure the controls
Evaluate unauthorized-action success, detection delay, false alarms, legitimate-task degradation, recovery time, rejected-command handling, and independent physical outcome checks.
Retain evidence
Record configuration, software and model versions, assumptions, test cases, exceptions, and reviewer approval. A lab or simulation test proves only the scope tested—not certification or safe deployment.
Before commissioning, conduct site-specific review with OT cybersecurity, control engineering, field operations, protection and safety personnel, IT, and the system integrator. Confirm the authority model, operating limits, fallback behavior, support responsibilities, and acceptance evidence. Reassess after material changes to equipment, firmware, network topology, AI models, vendors, or operational procedures.
How UtilityGrid frames secure Physical AI and AIoT workflows
UtilityGrid’s Physical AI and AIoT approach separates identification, sensing, AI-supported decisions, and authorized physical action. Security controls apply across each function, while the action path requires additional authority checks, equipment constraints, human oversight where required, and verification of the resulting state. The actual design depends on the utility’s assets, control architecture, interfaces, and operating procedures.
Related UtilityGrid resources
These resources provide additional context for action functions, control interfaces, system integration, pilot evaluation, and trusted industrial AI research.
