Cisco Research: Multi-Turn Conversations Bypass LLM Safety Guardrails (infosecurity-magazine.com)

• Multi-turn conversations bypass safety guardrails in all major LLMs.
• Techniques used: roleplay, ambiguity, reframing.
• Current single-prompt benchmarks underestimate risk.

"Cisco researchers tested major LLMs including ChatGPT, Claude, Gemini, Amazon Nova, and Grok and found that all could be tricked into bypassing safety guardrails through multi-turn conversations. Techniques included roleplay, ambiguity, and reframing requests after initial refusals. No model was completely safe. The study warns that current single-prompt benchmarks understate real-world risk, as attackers iterate across turns."