How do AI agents and robots change the threat model for power line repair?

Adding robots and AI agents to power line repair changes the threat model in three ways. Software now sits close to physical work, remote links carry commands to machines, and models turn sensor data into decisions. Treat each as a trust boundary, then apply industrial security standards to it.

What actually changes?

A crew on a pole or a drone inspecting a span used to depend on people, paper procedures and a radio. A robot adds firmware, a controller, a network link and an update channel. An AI agent adds a model that reads images or telemetry and suggests or takes an action.

None of that is exotic. Each piece is a familiar asset class. The shift is that a software fault can now have a physical consequence close to energised equipment, so the cost of a wrong input is higher.

Where are the new trust boundaries?

Think in four places.

The command path. Who can tell the machine to move, and how is that proven? A remote operator console, a fleet service and a mobile app are all ways in.

The data path. Cameras, lidar and thermal sensors feed a model. If an attacker or a fault alters that input, the model's output changes with it. Sensor spoofing and poisoned training data are the two cases to name in a risk assessment.

The update path. Firmware and model files arrive from somewhere. If you cannot say who built them and whether they were altered, you cannot trust them near a live line.

The supply chain. Open source libraries, vendor SDKs and pretrained models sit inside the machine. You inherit their vulnerabilities.

Which standards apply?

You do not need to invent a framework. Several exist.

The ISA/IEC 62443 series covers security for industrial automation and control systems. ISA describes a shared responsibility model with four roles: asset owners, integrators, product suppliers and service suppliers. That split is useful here, because a utility, a robot maker and an integrator each own a different slice of the risk (ISA).

For the robot itself, one industry explainer reports that the 2025 edition of ISO 10218 adds cybersecurity requirements, including a cybersecurity risk assessment, and refers to IEC 62443 (Fertigungstechnik). Read the standard text itself before you put clause numbers in a procurement document, since we only confirmed that summary.

For AI behaviour, the NIST AI Risk Management Framework gives a vocabulary for govern, map, measure and manage activities (NIST).

What should an AI agent be allowed to do?

Separate advice from action. An agent that flags a damaged insulator in an image is low risk. An agent that commands a tool to move near a conductor is not.

A sound pattern is to keep the model on the advisory side and put a deterministic, separately reviewed safety layer between any model output and any actuator. Humans approve physical actions until you have evidence, not enthusiasm, that you can relax that.

What to do

1. List every software component in the machine and fleet service, including models and their sources. An SBOM is the starting point.
2. Draw the command, data, update and supply chain paths and mark who controls each.
3. Require signed firmware and model updates, and verify the signature on the device.
4. Run a cybersecurity risk assessment that covers sensor spoofing and loss of the network link, and define a safe state for each.
5. Keep model output advisory for physical motion, behind a separate safety function.
6. Ask each supplier which IEC 62443 requirements their product meets, and ask for evidence you can read.
7. Scan the open source dependencies of any control software you own, and track new advisories for them.

Quick check

Sources

• ISA/IEC 62443 series overview
• NIST AI Risk Management Framework
• Fertigungstechnik: More Cyber Security in Robotics

Last reviewed: 18 October 2026.