Could the next cyberattack on a nation’s power grid be orchestrated by artificial intelligence? According to Demis Hassabis, CEO of Google DeepMind, that possibility is no longer a theoretical one. Speaking at the AI+ Summit in San Francisco, Hassabis said AI-enabled assaults on critical infrastructure-such as energy and water systems-are “probably almost already happening now,” even if current attacks are not yet leveraging the most sophisticated AI capabilities. His assessment underlines a rapidly shifting threat landscape where malicious actors exploit emerging AI tools to compromise essential services.
The dynamics of such attacks are rapidly changing. AI-powered cyber intrusions can unleash adaptive malware with the ability to change its signature to avoid detection, automate reconnaissance to find vulnerabilities in operational technology, and conduct precision strikes on control systems. In energy, the attackers can manipulate smart inverters or grid management software to disrupt supply or generate cascading failures. U.S. cybersecurity agencies warn that integrating AI into OT environments requires stringent measures, including human-in-the-loop supervision, failsafe mechanisms, and continuous validation of the AI models against regulatory and safety standards.
Hassabis’ warning comes against the backdrop of accelerating AI progress toward artificial general intelligence, or AGI. Though he had projected that AGI will be realized by 2030, recent breakthroughs in reinforcement learning and reasoning models have led other industry leaders to suggest timelines as short as three to five years. AGI-defined here as AI systems matching or surpassing human capabilities across most domains-would dramatically expand both beneficial and harmful potential. In adversary hands, AGI could autonomously plan and execute multistage cyber-physical attacks that combine digital intrusion with physical disruption in ways that outpace human response.
Those are not theoretical risks: AI-enhanced cyberattacks already account for an estimated 40% of global incidents, with capabilities ranging from crafting grammatically flawless phishing lures to poisoning AI models themselves. In critical infrastructure, this might mean corrupting predictive maintenance algorithms or load forecasting systems, making operators make decisions based on falsified data. Sandia National Laboratories has demonstrated that brain-inspired autoencoder neural networks can detect both cyber and physical anomalies in grid operations by fusing high-frequency physical measurements with sporadic cyber traffic data. Such systems can run on inexpensive single-board computers, making them deployable across legacy and modern equipment alike.
The notion of “p(doom)” the probability of catastrophic AI misuse-has gained currency as a way for the expert community to think about policy questions. Hassabis sets his own p(doom) at “non-zero,” consistent with survey data indicating that 77% of AI professionals believe technical researchers should be concerned about catastrophic risks. Recent studies, however, indicate that many AI experts are unfamiliar with basic safety concepts, such as instrumental convergence, the idea that advanced AI agents will have self-preservation subgoals regardless of their goals. These knowledge gaps create a risk of underestimating the dangers, especially among those who conceptualize AI solely as a tool that can be controlled.
Along with these threats, defensive strategies are emerging. Joint guidance from CISA, FBI, NSA, and allied agencies outlines principles for safely integrating AI into critical infrastructure: perform thorough risk assessments prior to deployment, establish protocols for governing AI models, and update incident response plans to account for AI-specific vulnerabilities. In the energy sector, AI can be both a target and a shield optimizing grid operations while opening attack surfaces. Safeguarding these systems requires robust testing environments, digital twins for safe experimentation, and industrywide standards for evaluating AI reliability.
The geopolitical dimension adds urgency: China’s push for a World AI Cooperation Organization reflects a bid to shape global governance, positioning its AI offerings as more accessible to developing nations. Meanwhile, the U.S. has pledged to “do whatever it takes” to maintain AI leadership, including countering Chinese influence in setting international norms. In this competitive climate, securing AI systems against misuse is not only a technical challenge but also a strategic one. The message from Hassabis is clear: the convergence of AI capability growth and critical infrastructure vulnerability demands immediate, coordinated action. As AI systems edge closer to AGI, the window for implementing resilient safeguards is narrowing, and the stakes-measured in the stability of power grids, water supplies, and other lifelines-could not be higher.
