Research · curated 29 Jul 2026
Jailbreaking Large Language Models via Multi-Task Embedding-based Prompt | Proceedings of the 2026 IEEE/ACM Third International Conference on AI Foundation Models and Software Engineering
First reported acm.org
Coverage timeline
Single-source research — first reported, latest, and curated coincide.
Why it matters
MTEA demonstrates a highly effective, low-cost jailbreak technique that reportedly bypasses leading commercial LLMs and existing defenses, showing that multi-task obfuscation can reliably break safety alignment.
Researchers present the Multi-Task Embedding-based Attack (MTEA), a jailbreak that embeds malicious instructions across three concurrent tasks (code understanding, language translation, and pattern adherence) to disrupt LLM safety alignment. On the AdvBench benchmark against six models including GPT-4o and Gemini-2.5-pro, MTEA reports a 100% attack success rate, defeats Perplexity Filter and SmoothLLM defenses, and cuts query costs by 90% versus baselines.