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

Researchers at Cisco have discovered that safety guardrails in major large language models — including OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini — can be bypassed through multi-turn conversation manipulation. The findings were reported by Infosecurity Magazine. The technique involves engaging the model in a series of exchanges that gradually steer it toward generating prohibited content, circumventing single-turn safeguards.

The Cisco researchers demonstrated that by carefully constructing a sequence of prompts, they could cause the models to produce responses that violate the models' stated safety policies. The vulnerability was found across multiple leading LLMs, indicating a systemic issue rather than a flaw in any specific model's implementation. The exact prompts used in the research have not been publicly disclosed.

The discovery highlights ongoing challenges in AI safety research, as current guardrails designed to prevent misuse can be circumvented with sufficient conversational skill. The affected companies have not yet publicly responded to the findings. The research underscores the need for more robust safety mechanisms that are resilient to multi-turn attacks, as these models continue to be deployed in customer-facing applications.