AI Model Claude Recreates PLC Exploit in Hours, but Hardware Pays the Price

Researchers at Forescout have demonstrated that the generative AI system Claude can translate a known programmable logic controller (PLC) vulnerability into a working exploit, doing so in eight hours and at a cost of $536. The experiment, first reported by Security Affairs, was intended to answer a lingering question in the industrial‑security community: whether artificial intelligence can autonomously adapt existing code to target new hardware platforms.

The team fed Claude the details of a previously disclosed PLC exploit and asked it to produce a version that would run on a different controller model. Within the allotted time, the AI generated a functional payload that, when deployed, successfully compromised the target device. The researchers note that the expense reflects the compute resources required for the AI’s inference and the cloud services used during the test.

While the initial success underscored the growing capability of large language models to assist in cyber‑offensive tasks, the follow‑up attempt to refine the payload had unintended consequences. An AI‑generated variant, meant to be more reliable, inadvertently caused physical damage to the PLC hardware, rendering the unit inoperable. The incident highlights a new risk vector: AI‑crafted code that can not only breach digital defenses but also trigger destructive outcomes in the physical world.

Industrial control systems have long been a focal point for security researchers due to their critical role in manufacturing, energy, and infrastructure. Traditional attacks on PLCs often require deep domain expertise and extensive manual reverse‑engineering. The ability of an AI model to shorten that timeline raises concerns about the accessibility of sophisticated exploit techniques to less skilled actors.

Forescout’s findings arrive amid broader debates about regulating AI tools that can be weaponized. Some experts argue for tighter controls on model access and output monitoring, while others caution that overly restrictive measures could stifle legitimate research and defensive innovation. The balance between fostering security advancements and preventing misuse remains a contentious policy challenge.

Looking ahead, the researchers plan to expand their testing to additional PLC families and to evaluate defensive measures that can detect AI‑generated malicious code. Their work serves as a cautionary example for operators of critical infrastructure, underscoring the need for robust monitoring, firmware integrity checks, and updated incident‑response strategies in an era where AI can accelerate both attack and defense.