Vedere Labs showed that Claude, an AI model, can port a remote-code-execution exploit from one WAGO programmable logic controller (PLC) to another. Human researchers still guided the process, proving that artificial intelligence is becoming a practical tool for industrial cyber-attacks. The proof-of-concept demonstrated a working exploit against a buffer-overflow flaw in the Nucleus FTP server.

Why this matters

Industrial control systems run factories, power plants and water-treatment facilities. A vulnerability that lets an attacker execute code on a PLC can sabotage processes, steal data or install persistent backdoors. By automating exploit adaptation, AI lowers the expertise needed to weaponize such flaws, expanding the pool of potential attackers.

How the AI did it

The Vedere team paired Claude with the open-source reverse-engineering suite Ghidra. They prompted the AI to analyze the original exploit’s shellcode—a tiny program that runs after the buffer overflow—and rewrite it to fit the firmware constraints of a different PLC model. Claude:

  • Modified the shellcode to bypass the target’s memory layout.
  • Achieved code execution that sent crafted network packets, proving the exploit worked end-to-end.
  • Tried to embed a command-and-control implant; that run permanently bricked the PLC, underscoring the risk of trial-and-error at scale.

Researchers still steered the process, verified results and dealt with the hardware failure. The experiment also revealed the high cost of API calls required for analysis and code generation.

The broader alarm

Both the Cybersecurity and Infrastructure Security Agency (CISA) and the FBI have warned that AI-generated scripts make industrial control systems an increasingly attractive target. The main concern is that the barrier to entry for sophisticated attacks is dropping; actors without deep reverse-engineering skills can now rely on AI assistants to craft functional exploits.

Limits and counter-points

The experiment is not a turnkey weapon. It required a skilled team to set up prompts, integrate Ghidra and interpret the output. The accidental destruction of a PLC shows the process remains fragile and error-prone. The expense of consuming large-language-model APIs may deter casual misuse.

What to watch

  • Continued advisories from government cyber-security agencies as AI tools become more accessible.

The Vedere Labs case makes clear that AI is no longer a theoretical aid for attackers—it can actively rewrite malicious code for new hardware, reshaping the threat profile of critical industrial systems.