New “Echo Chamber” jailbreak exposes how AI can be manipulated despite built-in safeguards.

Researchers at NeuralTrust have identified a method called Echo Chamber that uses innocuous prompts to steer AI models toward restricted content without triggering safety systems. By subtly altering conversation context over multiple turns, the technique avoids prohibited keywords while achieving harmful outcomes. Tests showed high success rates in generating disallowed responses, raising concerns about the ease and effectiveness of such attacks despite existing guardrails.

Read more:

https://www.securityweek.com/new-echo-chamber-jailbreak-bypasses-ai-guardrails-with-ease/