Eight Hours to Weaponize a Five-Year-Old Bug: What AI-Ported Exploits Mean for Maritime OT
When an AI Ports a Controller Exploit in an Afternoon, the Old Excuse for Ignoring Old Bugs Stops Working
September 4, 2026 | By The CyberSurv Team | CyberSurv
For years, defenders leaned on a comforting phrase whenever a nasty industrial vulnerability surfaced. “Too hard to weaponize.” The flaw was real, sure, but turning it into a working attack against your specific controller demanded firmware reverse-engineering, ARM shellcode wizardry, and a rack of test gear most criminals never bother to assemble. That friction was a kind of moat. On September 1, researchers at Forescout’s Vedere Labs published an experiment that shows the moat is draining.
They handed a frontier AI model a real programmable logic controller and asked it to do a job normally reserved for specialist exploit developers. It did the job. Not instantly, not cheaply, and not without a lot of human hand-holding — but it did it. And the moment that happens once, the economics of who can attack your operational technology quietly shift under your feet.
What the Researchers Actually Did
Vedere Labs already owned a working remote-code-execution exploit for a WAGO 750-852 controller. Their question was narrow and practical: could an AI take that exploit and port it to a different model, the WAGO 750-831, running firmware V01.04.16? Porting sounds like copy-paste. It isn’t. Memory layouts differ, function addresses move, buffer behavior changes. This is the grinding, expert work that keeps most vulnerabilities theoretical.
The underlying flaw is CVE-2021-31886, a pre-authentication buffer overflow in the Nucleus FTP server’s handling of the USER command. It carries a CVSS score of 9.8 and sits wide open on TCP port 21 before any login. A five-year-old bug, in other words. Exactly the kind teams triage to the bottom of the pile.
Working with Claude models and giving them terminal access, Ghidra, and a live target, the researchers watched the AI confirm the vulnerability, chew through the firmware, and hit the same dead ends a human would. At one point normal FTP processing kept zeroing out the payload buffer. The model figured out the workaround on its own — dropping the line terminators and swapping the command sequence to preserve the shellcode in memory. That’s not pattern matching. That’s problem solving.
The Numbers That Should Get Your Attention
Here’s where it gets interesting, and where the hype needs a cold splash of context. Reaching reliable code execution took eight and a half hours of sustained work and burned through $535.74 in API costs. A seasoned exploit developer would likely have moved faster and spent less. So no, the machines haven’t replaced the specialists. Not yet.
But look at what happened after the hard part was done. Once the AI had a foothold, it generated two fresh network payloads — an ICMP beacon and a UDP beacon that literally sends “PWNED” packets — in twelve minutes. Twelve. The Forescout team put it plainly: the first problem still takes expertise and time, yet adapting what comes next is becoming trivial to automate. That asymmetry is the whole story. Break in once, and the AI scales the rest.
There’s a sobering footnote, too. When the model tried to build a more advanced implant, one of its experimental payloads wrote to a flash-mapped memory region and permanently bricked the controller. In a lab, that’s an annoyance. On a live vessel or at a water plant, an autonomous tool blundering into physical memory isn’t a bug report — it’s an outage, maybe a safety event.
Why This Lands Hard on OT and Maritime
Ships, ports, and terminals run on exactly this class of device. PLCs govern ballast systems, cargo cranes, fuel management, and engine-room automation. Many of them were commissioned years ago, patched rarely, and connected to networks nobody fully mapped. The 750-series controllers in the experiment are cousins of gear humming away across the maritime transportation system right now.
Regulators already see the trend line. This week the UK’s NCSC warned that threat actors are increasingly targeting internet-exposed OT and edge devices and causing real disruption. And back in August, five U.S. agencies — NSA, CISA, FBI, DOE, and EPA — issued advisory AA26-231A, formally calling AI-generated exploit scripts aimed at Siemens controllers an active threat, a story we covered when it broke. The Forescout work isn’t an outlier. It’s a data point in a pattern that’s hardening fast.
Put those pieces together and the risk math flips. When exploitation was expensive, an obscure five-year-old FTP bug on a controller two subnets deep felt safe enough to defer. When an AI can absorb that expense and then replicate the result across a whole fleet of similar devices, “deferred” starts to look like “exposed.”
Stop Ranking Bugs by How Hard They Look
The most useful takeaway from Vedere Labs is a change in mindset. Prioritize your vulnerabilities by reachability, not by perceived difficulty. If an attacker — human or machine — can touch a device over the network, assume the difficulty tax on exploiting it is trending toward zero. The question stops being “could someone realistically build this exploit?” and becomes “can anything reach this port at all?”
Concretely, that means a few things for operators:
- Find every exposed controller. You can’t defend an asset you haven’t inventoried. Internet-facing OT should be treated as compromised until proven otherwise.
- Segment ruthlessly. Both NCSC and CISA hammer the same point — cut OT off from IT and the public internet, and choke the paths an automated tool would use to pivot.
- Re-triage the “unexploitable” pile. Those low-priority, hard-to-weaponize findings deserve a second look under the new economics.
- Watch the boundary. Ransomware crews are already hunting the remote-access gear that fronts these networks, a threat we detailed earlier this year.
This is precisely the analysis our RiskGuard 365℠ program is built to run continuously. Rather than a once-a-year snapshot, it maps what’s actually reachable across your OT and IT estate, weighs each finding against how attackers — increasingly AI-assisted ones — really operate, and keeps that picture current as your environment and the threat landscape shift. For vessels and shore facilities, our Maritime Cybersecurity Assessment brings the same reachability-first lens to the controllers, networks, and remote links that keep operations moving. And when you want to know whether a determined adversary can genuinely chain these weaknesses together, our penetration testing team pressure-tests the assumptions before someone else does.
Final Thoughts
Let’s keep the panic in check. An AI still needed most of a workday, real money, and a patient expert steering it to pull this off — and it torched a controller along the way. We are not at push-button industrial warfare. But the trajectory is unmistakable, and trajectories are what defenders get paid to read. The barrier that made old OT bugs feel harmless is eroding, and the part that scales — adapting an exploit across many devices — is the part AI already does well. The organizations that thrive won’t be the ones that panicked. They’ll be the ones that quietly stopped grading their vulnerabilities on a curve and started asking a simpler, sharper question: what can be reached, and what happens if it is?
Let’s Explore What’s Possible
Wondering which of your controllers an AI-assisted attacker could actually reach — and what it would cost you if one did? That’s a conversation worth having before the trajectory catches up. Our team can walk your OT and maritime environments through a reachability-first assessment and show you where to spend your next defensive dollar.
