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

Celý článek

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.