Researchers at Forescout’s Vedere Labs recently leveraged Anthropic’s AI tool, Claude, to migrate a remote code execution (RCE) exploit between two models of WAGO programmable logic controllers (PLCs). The process, while ultimately successful, demanded significant oversight, several hours of labor, and incurred hundreds of dollars in API fees.
Background on the Experiment
The team began with an exploit crafted for the WAGO 750-852 PLC, which exploited CVE-2021-31886. This vulnerability is a pre-authentication buffer overflow in the Nucleus FTP server, enabling unauthorized attackers to execute arbitrary ARM shellcode on the affected PLC. The goal was to adapt this exploit for the WAGO 750-831 model, evaluating whether AI could advance it to a full command-and-control implant, particularly after recent PLC attacks in the water sector.
Technical Challenges and AI Intervention
Claude Code, the AI tool, was equipped with access to a terminal, reference files, the reverse-engineering tool Ghidra, and the PLC itself. It confirmed the vulnerability using live probing combined with static firmware analysis, eventually generating a payload that caused the PLC to crash. This confirmed the flaw’s presence, but converting it into a controlled code execution proved to be complex.
Initially, the AI chased misleading paths, necessitating researcher intervention to adjust its analysis. Progress was stagnant until the team switched AI versions from Claude Sonnet 4.6 to Claude Opus 4.6. This version was instructed to seek assistance when uncertain about firmware specifics. A significant breakthrough was achieved when the AI identified and corrected why the injected code was being erased.
Outcome and Future Implications
Following this adjustment, Claude rapidly produced two functional payloads within 12 minutes. Researchers observed that although the initial execution was labor-intensive, subsequent iterations were swift, highlighting the scalability potential of this approach.
A subsequent attempt to develop a command-and-control implant faced obstacles. During this process, one payload inadvertently wrote to a flash memory region, permanently damaging the PLC. The financial aspect was notable, with the final RCE development phase alone exceeding $500 in API costs over more than eight hours.
Forescout remarked that while a human researcher might achieve results faster and more economically without AI, the potential reduction in expert intervention needed over time could allow AI to significantly lower marginal costs across numerous targets simultaneously.
This experiment underscores the evolving role of AI in cybersecurity, prompting questions about its future efficiency and cost-effectiveness in similar tasks.
