← all tags
// tag: ai-research
can you jailbreak an ai by making it act drunk
Yes. UNSW got five models drunk three ways — Mistral hit 90% attack success. The strange part: a real jailbreak defense tripled the attack instead of stopping it.
