By Marco C. Janssen, UTInnovation, the Netherlands
The question is no longer whether AI will be used to attack the power sector.
The power utility industry has spent decades building an increasingly sophisticated digital infrastructure to improve reliability, efficiency and operational performance. SCADA, EMS, DMS, ADMS, smart meters, IoT devices, cloud platforms and distributed energy resources have transformed the electricity system into one of the most connected pieces of critical infrastructure in the world. Now, however, artificial intelligence is changing the cybersecurity equation, and not necessarily in the utility’s favor.
Recent incidents should be regarded as a wake-up call. In July 2026, AI systems operated during security testing by Anthropic and OpenAI reportedly gained unintended access to real-world systems and demonstrated the ability to identify and exploit vulnerabilities. Anthropic subsequently disclosed that AI models had accessed the systems of three companies during testing. More concerning utilities, US authorities have recently warned that hackers are targeting specific programmable logic controllers used in energy, water and industrial infrastructure, with AI tools helping attackers develop exploitation scripts and reduce the technical expertise required to launch attacks.
This is fundamentally different from traditional cybercrime. Historically, an attacker needed considerable technical knowledge, time and resources to understand an industrial environment, identify vulnerabilities and develop an attack. AI can increasingly automate parts of this process. It can analyse documentation, identify weaknesses, generate code, adapt attack strategies and operate continuously. The result is potentially a dramatic reduction in the cost, time and expertise required to attack critical infrastructure.
For power utilities, the consequences are particularly serious because the IT and OT worlds are converging. An attack on an office network may result in stolen data or financial loss. An attack that reaches the OT environment can potentially affect breakers, protection systems, substations, generation assets or control systems. The boundary between a cyber incident and a physical incident therefore becomes dangerously thin.
The challenge is also not limited to attackers using AI against utilities. Utilities themselves are increasingly introducing AI into forecasting, predictive maintenance, asset management, network optimisation and control-room decision support. This creates a second category of risk because what happens when the AI itself becomes part of the operational attack surface? NERC has already identified AI-enabled attacks, data poisoning, malicious training inputs and over-reliance on AI outputs as emerging risks to electric-system reliability.
Imagine an AI-based operational system receiving manipulated sensor data indicating that a transmission line is overloaded. The AI recommends switching off the apparently overloaded line and redistribute power across the grid. An operator, accustomed to trusting the system, approves the recommendation. The resulting action may be technically valid according to the corrupted data, but operationally wrong. The attacker does not necessarily need to take control of the grid. It may be enough to make the utility’s own intelligent systems take the wrong decisions.
This creates a fundamental principle for the next generation of utility architecture as AI should never become an unquestioned authority in a safety-critical environment. NERC’s work on AI in real-time operations emphasises the importance of human oversight, validation and the ability for operators to question and override AI recommendations.
Utilities therefore should move beyond conventional cybersecurity and adopt an AI-resilient operating model. This means secure-by-design AI architectures, strict segregation between IT and OT, zero-trust access, continuous monitoring, independent validation of AI outputs, protected data pipelines, model governance, strong identity management and tested manual fallback procedures. On the upside AI can be used on the defensive side to detect anomalies, identify vulnerabilities, monitor behaviour and accelerate incident response.
The uncomfortable reality is that the power sector cannot simply choose not to use AI. The operational and economic benefits are expected to be enormous. But neither can utilities assume that existing cybersecurity frameworks are sufficient. The attack surface is evolving faster than many organizations’ security architectures.
The question is no longer whether AI will be used to attack the power sector. It is how quickly utilities can ensure that their AI-enabled grids are harder for AI-enabled attackers to compromise.
For an industry whose fundamental promise is reliability, resilience must now include not only the ability to withstand storms, equipment failures and human error but also machines attacking machines at machine speed. I think this is something to consider and act upon rather now than later or Skynet may become a reality sooner than we think.
Biography:
Marco C. Janssen is the CEO of UTInnovation and the former VP of Operational Excellence at TAQA, Digital Grid Leader for Latin America at EY and Director of the Smart Grid PMO at DEWA. He received his BSc degree in Electrical Engineering from the Polytechnic in Arnhem, Netherlands and has worked for over 33 years in the field of Power and Water O&M, Digital Transformation, Protection, AMI and Distribution and Substation Automation. He was a member of IEC TC57 WG 10, 17, 18, 19, the IEEE PES PSRC and CIGRE B5 and D2 WGs. He was the convenor of D2.35 and editor of the Quality Assurance Program for the Testing Subcommittee of the UCA International Users Group. He holds one patent, is the author of the book titled “Recreating the Power Grid”, has authored more than 53 papers, is co-author of 4 Cigre Technical Brochures and 2 books on SmartGrids and Electrical Power Substations Engineering and is the author of the “I Think” column in the PAC World magazine.
