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

Cisco researchers warn that many major large language models (including ChatGPT, Claude, Gemini, Nova and Grok) can have built-in safety guardrails bypassed through multi-turn, iterative conversations using techniques such as roleplay/persona adoption, ambiguity, reframing and escalation; no model was fully immune and single-prompt benchmarks understate real-world risk.