New Echo Chamber Attack Breaks AI Models Using Indirect Prompts

A groundbreaking AI jailbreak technique, dubbed the “Echo Chamber Attack,” has been uncovered by researchers at Neural Trust, exposing a critical vulnerability in the safety mechanisms of today’s most advanced large language models (LLMs).

Unlike traditional jailbreaks that rely on overtly adversarial prompts or character obfuscation, the Echo Chamber Attack leverages subtle, indirect cues and multi-turn reasoning to manipulate AI models into generating harmful or policy-violating content—all without ever issuing an explicitly dangerous prompt.

How the Echo Chamber Attack Works

The Echo Chamber Attack is a sophisticated form of “context poisoning.” Instead of asking the AI to perform a prohibited action directly, attackers introduce a series of benign-sounding prompts that gradually steer the model’s internal state toward unsafe territory.

Through a multi-stage process, the attacker plants “poisonous seeds”—harmless inputs that implicitly suggest a harmful goal.

Over several conversational turns, these seeds ...