Researchers Used Claude to Weaponize a PLC Exploit Across Device Models — And It Only Took Hours

The most telling story to cross the wire today wasn't a new zero-day or a fresh ransomware crew. It was an experiment. Researchers at Forescout's Vedere Labs took a known pre-authentication remote code execution flaw in WAGO programmable logic controllers, CVE-2021-31886, and used Claude to port the exploit from one PLC model to another. The result: arbitrary ARM shellcode executing on live hardware, achieved in hours and for a few hundred dollars. Three separate outlets covered this from different angles today — The Hacker News detailed the technical mechanics, Cybersecurity Dive framed it against the broader OT threat landscape, and SecurityWeek broke down the actual time and dollar cost of the experiment. That convergence matters, because this wasn't theoretical. It was a working demonstration of AI collapsing the labor cost of exploit development against industrial control systems.

What actually happened

The underlying vulnerability sits in the Nucleus FTP server's handling of the USER command — a classic stack-based buffer overflow that's been public since 2021. What's new is not the bug, it's the workflow. Porting an exploit from one PLC variant to another usually requires a security researcher with firmware reverse-engineering skills to manually adapt shellcode, offsets, and memory layout differences between models. Forescout instead handed the problem to Claude and had it do the adaptation work, ultimately getting arbitrary code execution on a different WAGO device than the one the original exploit targeted. SecurityWeek's writeup is blunt about the economics: hours of work, hundreds of dollars in compute. That's the number that should worry OT security teams — not because the vulnerability is new, but because the cost of re-weaponizing old, already-documented bugs against a fleet of similar devices just dropped dramatically.

Why this is the story, not just a story

OT and ICS environments are full of device families that share firmware lineage — same vendor, same protocol stack, slightly different hardware revisions. Historically, that diversity has been a mild form of protection: an exploit written for one model often needed nontrivial rework to hit a sibling model, and that rework was the bottleneck limiting how far an attack could spread. This experiment suggests that bottleneck is eroding. If a capable model can take a public CVE and adapt it across a product line in an afternoon, the population of exploitable devices tied to any single disclosed ICS vulnerability effectively grows overnight, without the attacker needing deep firmware expertise.

This isn't happening in a vacuum. Today's coverage also included the National Life Group CISO warning that AI-driven attacks could surface more vulnerabilities in the next six months than in the past thirty years of the industry combined, and separate reporting on OpenAI preparing to give partners early access to Astra, an AI model with what the company itself characterizes as critical cyber capabilities. The Financial Stability Board chair went so far as to call frontier AI cyber risk the most immediate concern to global financial stability. The WAGO experiment is a concrete, hands-on data point for all of that abstract anxiety — proof that the acceleration isn't hypothetical.

What defenders should take from this

The practical takeaway isn't panic, it's prioritization. Patch cycles for OT and ICS devices are notoriously slow, often measured in months or years rather than days, largely because of uptime requirements and change-control processes in industrial environments. That cadence assumed attackers needed real effort to weaponize old, known bugs against variant hardware. This experiment argues that assumption no longer holds for at least some vulnerability classes. Security teams running PLCs, RTUs, or any device family with known CVEs against sibling models should treat firmware-level exposure as a fleet-wide problem rather than a per-model problem, and push harder for compensating controls — network segmentation, strict access lists to management interfaces, out-of-band monitoring — as the interim layer while patch and replacement cycles catch up. CISA's own recent push to eliminate entire vulnerability classes at the design stage, rather than chasing individual CVEs, looks a lot more urgent in light of what a few hundred dollars of AI compute can now do to a decades-old bug.