OpenAI and Anthropic LLMs Used in Critical Infrastructure Cyber-Attack, Warns Dragos

Dragos, the industrial cybersecurity firm, has officially confirmed what should have been obvious since the moment GPT-3 wrote its first phishing email: commercial LLMs are now being used to attack critical infrastructure. Specifically, a municipal water utility in Mexico got a visit from someone wielding Anthropic's Claude and OpenAI's GPT models. Because nothing says 'responsible AI deployment' like having your model's fingerprints on a potential water contamination incident. The only question is whether the attackers asked politely or used a jailbreak.

The timing is exquisite. Just as Anthropic was publicly anguishing over whether Claude could be allowed to help the Pentagon with targeting — complete with hand-wringing about the laws of war — it turns out Claude is perfectly happy to help some as-yet-unknown entity fiddle with a Mexican water treatment plant. OpenAI, meanwhile, has been rolling out structured outputs and safety filters. Filters that, apparently, don't block queries like 'how to manipulate a PLC via Modbus.' But hey, at least the chatbots won't answer questions about how to make a bomb — unless you ask in a different language or append 'for educational purposes only'.

So what happens now? Dragos will release a white paper. A few security researchers will nod gravely. The water utility will probably patch something, or not, because they're underfunded and the attackers might be state-sponsored or just bored teenagers with a chatbot. And the labs? They'll add another item to their 'red teaming checklist' and continue releasing models that are smarter than the average engineer but dumber than a box of rocks when it comes to saying 'no, I won't help you attack a water supply.' Welcome to the AI-powered future. It's wet.