Single Line of Code Can jailbreak 11 AI models Including ChatGPT, Claude, and Gemini

Trend Micro researchers describe a black-box jailbreak technique called "sockpuppeting" that abuses assistant-prefill API functionality to inject compliant assistant-role messages and bypass LLM safety guardrails; the method required only a single line of code, produced functional malicious exploit code and leaked system prompts in successful tests across multiple models, and showed varying attack success rates (e.g., Gemini 2.5 Flash 15.7%, GPT-4o-mini 0.5%). The report highlights that defenses differ by platform — some providers block assistant prefills at the API layer while others rely on model-internal safety — and recommends message-ordering validation and adding these variants to AI red-teaming exercises.