Researchers Discover Major Security Gaps in LLM Guardrails

Unit 42 demonstrated that LLMs used as safety "AI Judges" can be manipulated by AdvJudge-Zero, an automated fuzzer that discovers low-perplexity token sequences which steer models to approve policy-violating content; the researchers reported a ~99% bypass success rate across common model architectures and recommend adversarial training to harden systems.