Recent advancements in large language models (LLMs) have significantly enhanced the ability of researchers to discover vulnerabilities and develop exploits. However, the complexity escalates when tackling specialized tasks such as reverse-engineering low-level firmware on embedded devices. This challenge is particularly pronounced in the realm of industrial control systems (ICS), which require expertise that typical AI models have yet to fully replicate.
Researchers at Forescout decided to investigate how far AI could go in mimicking the nuanced skills of human ICS vulnerability professionals. They employed AI tools to port a known exploit from one model of programmable logic controllers (PLCs) to another, a task that took 8.5 hours and still leaned heavily on input from human experts. While the AI didn’t fully replace the need for specialized knowledge, it did demonstrate a marked improvement in processing times and capabilities.
The outcome indicates that while AI can't yet transform ICS vulnerability discovery into a straightforward task for novice attackers, it definitely has potential to streamline the efforts of seasoned security professionals. The collaborative nature of AI research—where multiple agents can operate in parallel—could accelerate the process of developing new exploits. “AI has lowered the barrier to vulnerability research in higher-level software, and this experiment indicates that this trend is beginning to extend to low-level embedded systems,” Forescout researchers stated in their report.
Adapting Exploits to Other Devices
In their study, Forescout created a hypothetical attack scenario by adapting an exploit originally designed for one PLC model to another that was flagged as vulnerable, yet lacked existing exploit code. This scenario is particularly telling, as manufacturing systems often feature shared vulnerabilities across different models, yet a comprehensive assessment isn’t always conducted.
For instance, Forescout previously noted attempts to exploit a vulnerability in serial-to-IP converters from Lantronix (CVE-2025-67038). After initially patching only select models, the manufacturer later had to issue additional patches as exploitation attempts surfaced. This incident emphasizes the risk of incomplete patching across device offerings, enabling AI to assist adversaries in navigating unaddressed vulnerabilities effectively. “If manufacturers don’t perform thorough assessments, AI might empower attackers to adapt exploits to models that have yet to be patched,” said Daniel dos Santos, Forescout's VP of research.
In the Forescout study, the team specifically ported an exploit they had previously developed for CVE-2021-31886, demonstrating both the viability and challenges faced when shifting exploits across diverse models.
Shifting Attack Patterns
The evolving capabilities of AI are poised to alter the focus of attackers' strategies. While sophisticated vulnerabilities like remote code execution (RCE) in PLC firmware can lead to significant threats within operational technology (OT) networks, exploiting these vulnerabilities often requires a technical finesse that’s typically avoided in favor of easier targets like insecure protocols.
The introduction of AI tools changes the risk assessment landscape, potentially making it simpler to exploit previously “too difficult” vulnerabilities. “Threat actors often weigh the return on investment when developing exploits, and AI could fundamentally change that dynamic,” dos Santos explained. With AI's assistance, complexities could become manageable, allowing attackers to create deeper infiltrations and maintain persistence within systems.
The Future of AI in Cybersecurity
Forescout’s researchers utilized relatively older generation models—specifically, Claude Sonnet 4.6 and Claude Opus 4.6. However, the landscape has continued to evolve, with newer iterations, such as Anthropic’s Opus 4.8 and Fable 5.1, boasting enhanced capabilities in security research and exploit development. These advanced models could yield superior outcomes in vulnerability research than their predecessors.
In operational technology environments, the complexities of patching are compounded due to the imperative of maintaining critical processes. Asset owners often prioritize patches based on exploitability, but AI is creating a necessity to revisit these assessments. Organizations should account for potential AI-enhanced attack vectors in their risk management strategies.
Moreover, it could be crucial for organizations to revise incident response plans to include scenarios where AI enables rapid adaptation of exploits. This includes the need for regular updates to tabletop exercises and technical playbooks that cover potential exploit porting, lateral movement through systems, and stability issues from exploitation failures.
One notable caution from Forescout's research is the unpredictability of AI-driven exploitation. During their tests, an AI agent inadvertently compromised a device by attempting to deploy a malicious payload, resulting in a bricked system. Such unanticipated outcomes underscore the risks of integrating AI into exploit processes, emphasizing the consequences of miscalibrated actions in cyber-physical systems.
“The experiment highlighted the unforgiving nature of low-level embedded system exploitation,” the researchers noted. As AI agents become more autonomous, the repercussions of flawed decision-making in exploitation scenarios could extend well beyond mere digital misconduct, threatening operational safety.
