Researcher Exposes Systemic LLM Safety Flaws Across Major AI Models

Security researcher Dave Kuszmar discovered multiple jailbreak techniques that bypass safety guardrails in all major LLMs, obtaining instructions for making weapons and drugs. The exploits exploit the same restrictions meant to secure models, and Kuszmar reports AI labs have been unresponsive to vulnerability disclosures. He calls for slowing LLM deployment and increasing transparency before wider integration into society.