US DOE flags AI energy security risks, urges industry to tread carefully

Artificial intelligence could help the US optimize its energy supplies and manage its power grid, but the emerging technology could be also be used to sabotage critical infrastructure or have unintended consequences, the US Energy Department warned.

The federal agency on April 29 released interim risk assessment on AI in the energy sector. It identified ways AI could improve the nation's energy security, but it also highlighted the potential for the technology to be a weapon in the wrong hands, and even well-intended uses could create new attack vulnerabilities or lead decision-makers astray, the DOE said.

The assessment "reveals the clear need for regularly updated, risk-aware best practice guidance to facilitate the safe, secure, and beneficial deployment of AI in critical energy infrastructure," the DOE's Office of Cybersecurity, Energy Security, and Emergency Response (CESER) said in a news release.

CESER plans to publish an updated assessment later this year after gathering input from energy sector stakeholders on opportunities for AI and knowledge gaps in the space.

Double-edged sword

According to the interim assessment, AI can make the energy industry more secure by helping system operators parse large volumes of data, simulating weather events, predicting maintenance needs and detecting anomalies or threats, among other applications. The assessment was one of several studies the DOE published on April 29 about the technology's implications for the energy sector, after President Joe Biden in 2023 directed federal agencies to research the security and safety risks.

But the use of faulty AI models to inform decision-making or to automatically execute decisions could lead to bad outcomes, CESER said in the report. For example, an AI model may place undue weight on economic gain over grid reliability. "This risk is especially pronounced when energy systems are under stress, as during extreme weather events," the office said.

Systems based on machine learning that make predictions by processing large datasets are prone to attacks, CESER said. For example, adversaries may "poison" the AI model by inputting faulty data, or extract data from the model to access sensitive information about an energy system.

AI can be used to launch a cyberattack on energy infrastructure or to plan a physical attack. "Automatic parsing of text for vulnerabilities can reduce the amount of time and manpower an adversary needs in order to conduct reconnaissance on energy infrastructure being targeted for a potential attack," CESER said.

Another risk comes in pairing AI with other technologies, such as unmanned drone systems that could be used to attack energy infrastructure, CESER said.

However, while AI is risky if "deployed naïvely," the DOE office said human supervision, the protection of important models or datasets, and other best practices can help mitigate these risks.

"Regardless of whether AI technologies are deployed in service of energy infrastructure, the potential remains that hostile actors who wish harm to the sector may find ways to leverage AI to their own ends," CESER said.

Other studies

A second report published by the DOE on April 29 examined near-term opportunities for AI to improve the planning, permitting and operation of grid infrastructure. A third report, released by the DOE's Argonne National Laboratory, identified long-term energy sector challenges that could potentially be solved by AI.

Separately, the DOE's Lawrence Berkeley National Laboratory has been studying the electricity needs of AI and other types of computing. According to the lab's preliminary report, at least half of the growth in power demand from US datacenters since 2016 has come from AI.