AI's Role in Vulnerability Research for Industrial Control Systems
In recent months, large language models (LLMs) have made notable strides in their ability to assist with vulnerability research and exploit development. Despite these advancements, the complexities involved in identifying vulnerabilities in low-level firmware associated with specialized embedded devices remain a significant challenge. Researchers from Forescout, a company specializing in industrial IoT security, sought to explore the extent to which AI models can replicate the expertise traditionally held by human ICS vulnerability researchers.
Their investigations centered around using AI to adapt an existing exploit from one type of programmable logic controller (PLC) to another, which showcased a mix of promise and current limitations. The process, which took a total of 8.5 hours, still relied heavily on guidance from human experts. This suggests that while AI can aid in exploit development, it hasn't yet simplified the process to a level where untrained individuals can easily exploit vulnerabilities.
The Forescout team noted, “AI has already lowered the barrier to vulnerability research and exploit development in higher-level software. This experiment suggests that the same progression is beginning to reach low-level embedded systems, although substantial barriers remain.” This highlights the ongoing evolution in AI capabilities within security research, though seasoned hackers will still need to invest time and effort when leveraging these tools.
Porting Exploits Across Device Models
Forescout's researchers undertook a tailored test to mimic a real-world attack scenario: the task at hand was to adapt a proof-of-concept exploit for one PLC model to another model sharing the same vulnerability but lacking an existing exploit. This scenario is particularly significant since different models from the same vendor might have similar vulnerabilities, despite differences in firmware and architecture. Often, manufacturers do not conduct thorough assessments across their product lines after patching a known flaw, leading to potential security oversights.
A case in point is a previously reported vulnerability (CVE-2025-67038) affecting serial-to-IP converters by Lantronix, where patches were only issued for select device series. Later, after exploitation attempts were detected, the company moved to include additional product models. “I do believe that in case manufacturers do not perform a comprehensive assessment of models affected by a vulnerability, AI can now help attackers to do it and to port exploits to models that may not have been patched,” asserts Daniel dos Santos, VP of research at Forescout.
For their experiments, the researchers successfully ported an exploit related to CVE-2021-31886 from one Wago PLC to another that had already received a patch.
Shifting Attack Strategies
The growing proficiency of AI in achieving tasks associated with ICS vulnerability exploitation is poised to reshape how attackers select their targets. While remote code execution (RCE) vulnerabilities in PLC firmware present significant risks, the intricacies involved typically steer many real-world attackers towards easier targets, like insecure protocols or vulnerabilities in human-machine interfaces (HMIs).
“The biggest risk is that vulnerabilities that are considered ‘too difficult to exploit’ will become easier to exploit,” dos Santos explains. “Threat actors have a sort of ROI calculation when spending time to develop an exploit, and something that helps them create or port exploits changes this calculation: […] RCE can give very granular control to attackers to create persistent implants, execute low-level lateral movement, and so on.” The implications of AI's assistance in this arena could lead to more complex and damaging exploits than previously deemed possible.
The Future of Exploitation Models
The Forescout researchers utilized older-generation models during their research, such as Claude Sonnet 4.6 and Claude Opus 4.6, which were released earlier in the year. Since then, newer models, including Opus 4.8 and Fable 5, have emerged, boasting enhanced capabilities in security research. These versions, particularly Fable, are making a marked impact due to their advanced reasoning abilities, which are becoming crucial in vulnerability research.
As AI models continue to evolve, their effectiveness in different domains, including exploit development for embedded systems, will likely improve. This growth is essential, especially since patching in operational technology (OT) environments poses distinct challenges compared to traditional IT networks. Devices are integral to critical operations, and offline patching can result in significant downtime and operational risks.
With complex assessments now required for vulnerabilities, AI is set to reshape how organizations approach incident preparedness and vulnerability management. It’s vital for asset owners to consider AI-assisted attack vectors in their security strategies. Forescout recommends that security teams update their tabletop exercises and incident response playbooks, factoring in rapid exploit adaptation scenarios and lateral movement possibilities.
Interestingly, AI exploitation can introduce unpredictability, as demonstrated in Forescout’s study, where an AI agent attempted to deploy a malicious payload but inadvertently caused irreversible damage to the device by writing to a protected memory area. This incident highlights how unintended consequences can arise, emphasizing that as AI tools gain broader access to cyber-physical systems, mistakes can lead to severe repercussions.
Ultimately, this investigation illuminates both the potential and challenges that AI brings to ICS security. The evolving landscape demands that organizations proactively reassess their security strategies and remain vigilant in the face of advances in AI-assisted cyber threats.
