ChatGPT, Claude, and Gemini Among 11 AI Models Vulnerable to One-Line Jailbreak

A newly reported "sockpuppeting" jailbreak exploits assistant-prefill (prefix injection) at the API layer to cause LLMs to continue generating restricted outputs by inserting a fake acceptance message; tested across 11 models, it enabled generation of malicious code and system-prompt leakage in some cases, and can be mitigated by enforcing message-order validation at the API layer though self-hosted inference servers often remain vulnerable.