AI is getting closer to being able to exploit OT, and that's very bad news for critical infrastructure
Artificial intelligence has moved from theoretical labs into the control rooms of factories, proving it can adapt a remote‑code‑execution (RCE) flaw from one programmable logic controller (PLC) to another, a development that threatens the reliability of critical infrastructure.
AI Can Translate Exploits Across Closed‑Source PLCs
Forescout researchers demonstrated that an AI model, given sufficient guidance, was able to port an existing RCE vulnerability to a different PLC platform built on closed‑source firmware. The experiment succeeded in triggering a denial‑of‑service (DoS) condition and in executing attacker‑supplied ARM shellcode, showing that AI can bridge the gap between disparate industrial devices. AI can port RCE exploits to PLCs marks a technical milestone that blurs the line between software‑only attacks and physical‑process disruption.
The researchers began with a known vulnerability and used the AI to generate payloads tailored to the target PLC’s architecture. Because PLCs often lack the frequent patch cycles of IT systems, a successful exploit can linger undetected for months. This capability, however, required the AI to be steered by experts who understood both the original exploit and the target device’s internals.
While the proof‑of‑concept proves feasibility, it also highlights a new attack surface: any exposed PLC that communicates over network interfaces could become a candidate for AI‑assisted exploitation, provided an adversary can invest the necessary expertise.
Economic and Technical Barriers Keep Criminal Use Limited
The Forescout team reported that the final stage of developing a functional RCE cost more than $500 in API usage, a figure that dwarfs the budget of most cybercrime operations focused on quick returns. In addition to monetary cost, the process demanded extensive manual input from researchers to shape the AI’s output, a level of specialist knowledge that is scarce outside nation‑state labs.
When the team attempted to extend the exploit beyond the initial RCE, the target PLC was permanently bricked, demonstrating that the AI’s generated code can be unstable when pushed past a narrow scope. This instability raises the risk for attackers: a failed attempt could take critical equipment offline, drawing immediate attention and retaliation.
Because cheaper, more reliable attack vectors—such as phishing or ransomware—remain available, the cost‑benefit calculus currently disfavors AI‑driven OT attacks for most criminal groups. The high entry barrier acts as an inadvertent deterrent, buying time for defenders to improve monitoring and patching practices.
Nation‑State Interest Outpaces Criminal Incentives
The report cites the 2025 Sandworm operation against Poland’s power grid as a reminder that state actors possess both the resources and strategic motives to pursue sophisticated OT attacks. Unlike profit‑driven criminals, nation‑states can absorb the $500‑plus expense and the need for deep domain expertise, viewing disruption as a geopolitical lever.
Sandworm’s attack demonstrated how a coordinated campaign can exploit vulnerabilities in power‑generation equipment, causing widespread outages. The Forescout experiment shows that AI could accelerate the discovery and weaponization phases for such actors, potentially shortening the window between vulnerability identification and operational impact.
However, even nation‑states face practical limits: the bricking of a PLC during the experiment underscores the risk of collateral damage that could expose the operation. Strategic planners must weigh the benefits of a precise, AI‑crafted exploit against the possibility of unintended system failure that could undermine the intended political message.
What This Actually Means For You
- AI‑assisted exploits are technically possible but costly, so most attackers will stick to easier methods for now.
- Industrial environments with exposed PLCs should prioritize network segmentation to limit AI models from reaching device interfaces.
- Monitoring for unusual API traffic can reveal attempts to leverage costly AI services, offering an early warning sign.
- Nation‑state threats remain the primary driver of sophisticated OT attacks; organizations should align their risk assessments with geopolitical developments.
- Investing in regular firmware updates and vendor advisories reduces the pool of vulnerable PLCs that AI could target.
Immediate Action Steps
Begin by auditing all PLCs that are reachable from corporate or internet‑facing networks; isolate any that do not require remote access behind firewalls or air‑gapped segments. Next, establish a baseline of normal API usage patterns for any cloud services interacting with OT devices, and configure alerts for spikes that could indicate AI‑driven exploit attempts.
Finally, engage with PLC manufacturers to confirm you are subscribed to their security bulletins and apply any available patches promptly, even if the devices run closed‑source code. Keeping firmware current narrows the attack surface that AI can exploit.
Frequently Asked Questions
Can AI automatically find zero‑day flaws in PLCs?
The Forescout experiment shows AI can adapt a known RCE to another PLC, but it required heavy researcher guidance and cost over $500 in API usage, indicating full automation of zero‑day discovery is not yet practical.
Why are nation‑state actors more likely to use AI against OT?
State actors have the budget and expertise to absorb the high costs and specialist input needed for AI‑assisted exploitation, as illustrated by Sandworm’s 2025 power‑grid attack.
What immediate defenses can protect my industrial network from AI‑driven attacks?
Segregating PLCs from external networks, monitoring API usage for anomalies, and applying vendor firmware updates are the most effective steps highlighted by the researchers.
What Do You Think?
Given the high cost and specialist skill required, should industry focus on hardening OT now or wait for AI‑driven attacks to become cheaper and more accessible?