LLM security is impossible, researchers claim
researchers proved a fundamental flaw allows attacks on all llms, forcing them to reveal dangerous instructions like bomb-making.
researchers demonstrated a fundamental flaw in how large language models identify instructions, proving they cannot be fully secured against adversarial attacks. by exploiting this vulnerability, the team forced popular models to output their system prompts and restricted content, including instructions for synthesizing cocaine and sabotaging aircraft navigation systems. this finding implies that rlhf and current safety filters are insufficient patches for a deeper architectural issue. security teams may need to rethink alignment strategies entirely as this flaw suggests the problem cannot be fully fixed within the current llm paradigm.