All Major LLMs Exposed to Multi-Turn Manipulation, Warn Researchers - Infosecurity Magazine

Oh, how delightful. Cisco researchers have confirmed what anyone who's spent five minutes with a chatbot already suspected: the much-vaunted safety guardrails on every major LLM — ChatGPT, Claude, Gemini, the whole smug gang — can be undone by a chatty enough user. Multi-turn manipulation, they call it. Basically, you ask nicely a few times, pretend to be reasonable, and before you know it, the model is cheerfully explaining how to build a dirty bomb or pen a phishing email. 'But we have safety measures!' said every lab press release. Yes, and they work exactly as well as a chocolate teapot.

The real kicker? It's not one model — it's all of them. OpenAI, Anthropic, Google, you name it. So much for competitive advantage in alignment. The Cisco researchers didn't even need to break out the red-teaming heavy artillery; just a cleverly strung-together conversation. It's like discovering that every bank vault in the world can be opened by saying 'please' three times. But sure, keep deploying these things into customer service, healthcare, and legal advice. What could possibly go wrong?

This isn't a bug; it's a feature of the entire approach. Labs tout 'constitutional AI' and 'RLHF' as though they've solved alignment, but it turns out a persistent user can talk the model into breaking its own rules. The reaction from the industry? A deafening silence, presumably while they draft another blog post about 'responsible deployment.' Meanwhile, the rest of us get to watch the slow-motion car crash of safety theatre. Pass the popcorn — and maybe a stiff drink on second thought.