ChatGPT safeguards can be hacked to access bioweapons instructions — despite past safety claims: report

To circumvent the models’ defenses, the publication employed a jailbreak prompt: a series of code words that hackers can use ...