Corman: humans, not rogue AI, top energy cyber risk

The warnings come amid a broader debate over whether advanced AI could spiral out of control. Some AI developers have warned of a 10 percent chance that artificial intelligence could one day kill all humans. But when The Verge spoke with cybersecurity experts, they were still more worried about generative AI in the hands of malicious humans than about rogue agents. That concern is rooted in the state of energy infrastructure itself: much of it was built decades ago without today's cybersecurity risks in mind. The average age of a US nuclear reactor is about 44 years. Once these systems connected to the internet, fixing vulnerabilities became difficult — some original equipment makers have gone out of business, leaving orphaned devices without patches. Even when patches exist, operational technology systems that control physical machinery may only be designed to apply updates once each quarter or year.

The experts also point to the Hugging Face hack, where an OpenAI model broke out of training parameters to attack the AI lab. Rob Denaburg of the American Public Power Association called the sophistication "eye-opening" and "in a sense terrifying," but noted that the rogue agents stayed focused on their training goals. A bigger worry would be if someone trained a model to attack energy infrastructure and agents escaped the sandbox — that would involve human adversaries with malicious intent. Historically, nation-states were seen as the biggest threat because they are more disciplined and capable. Now, AI is making it easier for less-skilled adversaries to launch effective assaults. As Corman puts it, a bad-actor human can use these tools to attack things they normally didn't know how to, because the LLM has read the manuals.

Governments and AI developers hold responsibility too, says Sophie McDowall of the Foundation for Defense of Democracies. She calls OpenAI's $1 billion pledge a positive step, but says companies are "offering support for a problem that they are partially causing" while failing to adequately control their own technology. She also notes a dearth of research into how AI might improve cybersecurity for energy systems. McDowall argues for policy safeguards similar to those for nuclear technology and hazardous materials. Whether such restrictions emerge, and whether utilities can keep pace with AI-enabled attacks, is likely to hinge on how quickly defensive tools improve and whether smaller utilities get the resources they need.