
Two chatbots walk into a wind farm and switch it off before anyone finishes their coffee
Ourtake
The real vulnerabilities were default passwords and misconfigured devices. The chatbots just found them faster than any human attacker.
What happened
Researchers at Shinobi Security tested two off-the-shelf LLMs, Anthropic's Claude and Moonshot's open-weight Kimi, against a replica of a wind farm attacked in Poland in December 2025. Starting from VPN credentials, both models shut the simulated farm's grid connection in under two hours, with one spoofing weather data to trigger safety shutdowns. The findings were due to be presented at the SANS CyberThreat Summit on 16 November.