New FATS attack exploits LLMs, highly susceptible GPT-4.1 and DeepSeek-R1

Researchers have developed a new prompt injection attack called FATS (Feign Agent Attack with Toxic-shots) that exploits vulnerabilities in large language models (LLMs). This attack method manipulates LLMs by obfuscating preference extraction and compromising toxicity samples, leading them to generate harmful outputs. Experiments show that prominent models like GPT-4.1 and DeepSeek-R1 are highly susceptible to FATS, highlighting the need for careful analysis of security-related training data to build more secure LLMs. AI

IMPACT Highlights a new class of vulnerabilities in LLMs, potentially impacting their safe deployment and requiring new defense mechanisms.

RANK_REASON Research paper detailing a new attack method against LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• AdvBench
• DeepSeek-R1
• FAQuery
• GPT-4.1
• reinforcement learning from human feedback
• Yupeng Ren

• safety
• paper

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →