Robot Bouncers Wave Anyone In: Researchers Automate Jailbreaks for ChatGPT, Bard and Claude
Ourtake
Turns out the lock works fine until someone builds a machine that tries a million keys a minute.
What happened
Researchers found that automatically constructed adversarial attacks can trick major LLMs, including ChatGPT, Bard and Claude, into producing unintended and potentially harmful content. The attacks can be generated at scale and work on closed-source and public chatbots, and the researchers suggest they may not be fully preventable.