New Attack Exposes Stateless Multi-Turn Vulnerabilities in LLMs

A new multi-turn attack method called Transient Turn Injection (TTI) has been developed by researchers, which takes advantage of stateless moderation in large language models by spreading adversarial intent over separate interactions. Unlike traditional jailbreak techniques that depend on a continuous conversational context, TTI employs automated attacker agents driven by LLMs to systematically test and bypass policy enforcement. An assessment of leading models from OpenAI, Anthropic, Google Gemini, Meta, and notable open-source alternatives showed marked differences in their resilience, with only a few architectures demonstrating significant inherent robustness. The research paper can be found on arXiv with the reference number 2604.21860.