Researchers at ETH Zurich created a jailbreak attack that bypasses AI guardrails

Artificial intelligence models that rely on human feedback to ensure that their outputs are harmless and helpful may be universally vulnerable to so-called ‘poison’ attacks.