AI-Powered Hacking of Operational Technology: A Need for Human Guidance
In the contemporary landscape of cybersecurity, discussions surrounding the capability of artificial intelligence (AI) to autonomously execute operational technology (OT) hacks are garnering substantial attention. Recent studies reveal that, while fully AI-driven cyberattacks against OT systems remain a distant reality, there is still growing concern over the long-term implications as AI continues to evolve. Experts warn that for now, human involvement is necessary to navigate the complexities of such attacks.
Researchers from two distinct teams have employed varying methodologies and selected different models to investigate the capabilities of advanced AI. Their findings suggest that under human guidance, AI can autonomously execute each phase of a multi-stage attack on an unfamiliar OT network with relative efficiency. However, many of these attacks are described as "inefficient" and "noisy," indicating that while progress has been made, hurdles remain.
The expertise of the human operator plays a pivotal role in determining the success of these attacks. The ability to craft effective prompts and give valuable guidance significantly influences the outcome. A noteworthy observation from one of the studies, conducted by Booz Allen Hamilton (BAH), highlighted that once complex technical knowledge and specific skill sets were required to breach industrial control systems, these barriers have now been significantly reduced. As the report states: “Specialized OT knowledge, unfamiliar equipment, and complex control environments are no longer meaningful barriers to attack.”
Harry Thomas, co-founder and CTO of Frenos, emphasized the role of human oversight in the hacking decision-making process, suggesting that such involvement will persist “for the near future.” He noted that while current AI models show promise, they still need robust "agentic harnesses" to support their operations. Unattended AI attacks pose risks of detection, as even minor errors could easily alert security systems responsible for monitoring the OT environment. Thomas recounted his own experiences as a former hacker, saying that mistakes in an OT setup would often trigger automated response systems designed to identify and neutralize threats.
The talent required to effectively leverage these AI systems largely depends on the model being utilized. Some, like the open weight models, require higher levels of expertise, whereas frontier models have shown the capacity to produce results with comparatively less human input.
Rob Bair, a principal industry advisor specializing in cybersecurity at Anthropic, articulated concerns over the implications of democratized access to AI tools. “AI-enabled cyberattacks will likely put critical infrastructure at greater risk just because it lowers the bar to entry,” he stated, citing the continuous improvement of AI models. He cautioned that even a novice attacker could easily conduct research online about critical systems like water plants and manipulate AI to aid in crafting targeted attacks.
These sentiments were echoed by other experts who noted that as AI becomes increasingly proficient, the risk to essential service providers, such as water and power utilities, will also heighten. The CEO of Dragos, Rob Lee, asserted that adversaries equipped with AI will find it easier to identify and exploit vulnerabilities within vital infrastructures, demanding heightened vigilance among cybersecurity professionals.
While concerns over risks are valid, Marcus Sachs from the Center for Internet Security cautioned against overstating the threat posed by AI technology. He highlighted that the knowledge and skill barriers to hacking OT have historically not been insurmountable, especially for state-sponsored attackers. His perspective emphasizes that the coding languages and communication protocols specific to OT systems can be learned through determination and resourcefulness.
In a meticulously air-gapped laboratory designed to simulate an OT network environment, researchers at BAH assessed the potential of leading AI language models. The lab was structured to mirror systems commonly found in manufacturing environments. During tests, they utilized advanced AI models from renowned providers, albeit under strict restrictions to monitor and control any kinetic actions taken by the AI.
The results from BAH’s investigations revealed that both models managed to successfully navigate the entire array of eight test scenarios, demonstrating a remarkable level of reasoning and understanding of the OT environments they were probing. Additionally, the models were able to chain vulnerabilities into coherent attack paths, showcasing a degree of adaptability even when initial attempts failed.
Despite the findings underlining the models’ efficiency, experts like Adrian Sanabria urged caution, emphasizing that theoretical capabilities do not necessarily imply practical motivation for malicious actors. In practical terms, attackers typically prioritize easier, more cost-effective methods to achieve their objectives.
As the discussion around AI-driven cybersecurity evolves, it is clear that human oversight, expertise, and imagination will remain essential components in navigating the complexities associated with the integration of AI in hacking operational technology.