Forescout researchers utilized Anthropic's Claude AI to port a remote code execution (RCE) exploit between WAGO PLC models. While successful, the process highlighted the high cost of API usage and the critical need for human oversight.
- Forescout's Vedere Labs used Claude AI to adapt an RCE exploit from WAGO PLC model 750-852 to 750-831.
- The process required over 8 hours of work and cost more than $500 in API fees.
- While the AI eventually produced working payloads, one attempt permanently bricked the target device.
In a detailed cybersecurity experiment, researchers at Forescout’s Vedere Labs explored the capabilities of generative AI in the realm of industrial sabotage. By leveraging Anthropic’s Claude, the team attempted to port a working remote code execution (RCE) exploit across different models of WAGO programmable logic controllers (PLCs), simulating the types of threats currently facing critical infrastructure like the water sector.
The research focused on CVE-2021-31886, a pre-authentication buffer overflow vulnerability within the Nucleus FTP server. This specific flaw allows an unauthenticated attacker to execute arbitrary ARM shellcode on the target device. To facilitate the port, the researchers provided the AI with access to a terminal, Ghidra for reverse engineering, and the physical PLC hardware.
Why This Matters
BozokMedia analysis shows that the primary danger is not the AI's current autonomy, but its potential for scaling. While a human expert might currently port an exploit faster than an AI, the AI's ability to iterate rapidly once a breakthrough is made means that the marginal cost of attacking hundreds of similar targets simultaneously could drop to nearly zero in the near future.
The true risk lies in the reduction of the expert intervention threshold, enabling lower-skilled actors to execute high-impact attacks.
The path to success was not linear. Initially using Claude Sonnet 4.6, the AI frequently chased incorrect leads. The breakthrough occurred only after switching to Claude Opus 4.6 and implementing a feedback loop where the AI was encouraged to ask for clarification on firmware details. Once the AI identified why the injected code was being erased, it generated two functional payloads within just 12 minutes.
However, the experiment also demonstrated the volatility of AI-generated code. In a subsequent attempt to build a command-and-control (C2) implant, the AI produced a payload that wrote to the PLC's flash memory region, resulting in the permanent 'bricking' of the hardware.
| Metric | Manual Expert Porting | AI-Assisted Porting |
|---|---|---|
| Initial Analysis | Precise & Methodical | Fast but Prone to Errors |
| Iteration Speed | Slow (Manual coding) | Ultra-Fast (Seconds/Minutes) |
| Resource Cost | Human Labor Hours | High API Costs ($500+) |
Frequently Asked Questions
1. Can AI currently hack a PLC without human help?
No. This experiment proves that significant expert guidance is still required to steer the AI and provide the necessary technical context.
2. What was the financial cost of the AI experiment?
The final stage of RCE development alone cost over $500 in API usage fees over an eight-hour session.
