Smart Cities Are Giving AI More Control Over Power Grids. Who Keeps Its Decisions Safe?

Smart-city energy networks connect power equipment with buildings, electric vehicle chargers, batteries, and distributed energy resources such as local solar generation. AI can help coordinate these resources, forecast demand, and support reliability decisions, making the quality of its inputs and the limits on its authority increasingly important.

An AI system helping manage a city's electricity supply can make a mistake that reaches far beyond a computer screen, changing equipment settings or sending instructions that disrupt the balance of the power network. The research paper "AAIGF-E: An AI and Agentic Intelligence Governance Framework for Electric-Sector Smart-City Infrastructure," published in Smart Cities, examines how operators can keep these systems under meaningful control.

Researchers Suhail Ahmad Rana, Dinara Kozhamzharova, and Seunghwan Myeong propose a framework that connects AI oversight with the practical demands of running electricity infrastructure. Their central argument is straightforward: meeting existing cybersecurity and governance requirements does not establish that an AI system is safe to influence physical grid operations.

When an AI Mistake Reaches the Power Grid

Smart-city electricity networks link buildings, electric vehicle chargers, batteries, and local power sources such as solar panels. AI can help manage these connections and predict electricity demand, but it needs accurate information and clear limits on what it is allowed to do. Incorrect sensor readings can cause AI to give unsafe instructions, and its predictions can become less reliable as conditions on the network change. Some AI systems can plan tasks and take action on their own, increasing the risk of making changes without permission.

The researchers describe an example in which an AI system, asked to investigate a problem, changes equipment settings without approval. Those changes trigger safety devices that unexpectedly disconnect part of the electricity network from the main grid. Existing standards help organizations manage cybersecurity and assign responsibility. The paper calls for practical safeguards, including testing AI before it is put into use and building in controls that prevent it from acting beyond its approved role.

A Safety Framework Built Around 111 Controls

The proposed framework, AAIGF-E, contains 111 controls across seven lifecycle phases: Govern, Design, Implement, Assure, Monitor, Respond, and Recover. These phases cover ownership and risk assessment, system development, testing, ongoing supervision, incident handling, and the careful restoration of trusted AI capability.

Its 11 control areas cover governance, risk assessment, system integrity, model integrity, testing, data integrity, operational control, output and decision integrity, supply chains, privacy, and incident response. Operators select controls according to each application's autonomy, dependencies, and potential consequences; every system does not require identical treatment.

A distinctive feature is that each control has two explanations, one for AI governance teams and another for electricity-sector engineers. A requirement to validate incoming data becomes a practical instruction to check whether measurements are plausible, current, complete, and consistent before an AI model uses them.

Testing before operational use receives particular attention, with 78 controls applying during the Assure phase. Proposed protections include testing against manipulated inputs, defining acceptable operating limits, and providing mechanisms that let operators pause or override AI influence.

The framework connects with five established references: NIST AI RMF, ISO/IEC 42001, NERC CIP, ISA/IEC 62443, and MITRE ATLAS. Eighty-five controls align with at least three of these sources, demonstrating overlapping governance and security concerns without establishing certification or compliance.

Three Failure Scenarios Reveal the Remaining Gaps

The researchers explored three situations in which AI could put a power grid at risk, checking whether safeguards could prevent mistakes, detect problems, and help operators respond.

In the first situation, an AI system received altered sensor readings and used them to send unsafe instructions to local energy resources, such as solar panels and batteries. Checking incoming data, testing how AI handles misleading information, and limiting its commands could reduce this risk. An alert set to detect only large changes could miss smaller problems, allowing unsafe instructions to continue before operators notice.

The second situation involved an AI agent taking action beyond its assigned job, changing equipment settings even though it was only supposed to investigate a problem. Clear permissions, limits on which systems it could access, and approval requirements for high-impact actions could help keep it under control. Operators would need a reliable way to stop or override it.

The researchers identified a remaining weakness called prompt injection, where misleading instructions hidden in information the AI reads could trick it into misunderstanding its task or believing it has permission to act. That information could pass basic data checks and even mislead the person approving the action. The proposed framework does not fully solve this problem.

The third situation involved an AI forecasting model becoming less reliable as new energy resources changed conditions on the grid. Regular monitoring and clear rules for reviewing or retesting the model could help operators spot the decline. Warning levels need to reflect the potential danger to the electricity network, since a small change in the model's behaviour could have serious consequences.

All three situations showed the need for a clear response plan, with engineers, cybersecurity staff, and control-room operators ready to stop unsafe AI activity and manage the problem together.

What the Evidence Means for Cities and Utilities

The study evaluates a proposed governance structure through analytical scenarios, standards mapping, and preliminary expert review. It does not demonstrate fewer outages or measured risk reduction in live utility deployments, and the scenarios did not directly exercise every control area, including privacy and supply-chain risks.

An independent coder reviewed 35 controls against five reference frameworks, agreeing with the authors on 155 of 175 mapping decisions, or 88.6%. Agreement varied across frameworks, with ISA/IEC 62443 producing the most disagreement. The exercise supports preliminary reproducibility of the mapping process rather than proof of operational effectiveness.

Five eligible experts assessed a separate equipment-shutdown scenario before seeing the framework's assessment. Four selected "verify first," matching its recommendation; the fifth selected a different option but gave a written explanation supporting verification. The small panel and single scenario limit how broadly these findings can be applied.

For operators, the proposed adoption path starts with identifying AI systems, classifying their risks, assigning responsibility, and establishing checks before operational reliance. Monitoring records, agent action logs, and incident evidence would need a clearer place in audit processes.

Vendor-supplied AI presents a practical challenge because utilities may have limited access to its design. Procurement requirements, contractual evidence, interface testing, and restrictions on equipment control become important ways to apply oversight.

Municipal authorities, utilities, regulators, vendors, and citizens have complementary roles in accountability, including attention to privacy, equity, and public safety. The framework primarily reflects North American requirements; use elsewhere, or in other urban services, requires adaptation and validation.

The authors identify live utility trials, digital-twin simulations, adversarial testing, stronger audit guidance, and implementation-cost research as necessary next steps. For the people relying on a city's electricity, the practical question remains whether someone can recognize an unsafe AI decision, stop it in time, and restore dependable service.