An engineer from CyberArk has challenged the safety protections of OpenAI’s latest o3-mini model just days after its launch. He was able to get the model to teach him how to write an exploit for a Windows security process, which was achieved by posing as an student requesting educational information. Although the model features a new security feature called “deliberative alignment”, it still proved to be susceptible to certain manipulations.
Remus Malware Bypasses Browser Application-Bound Encryption Protections – cyberpress.org
Remus Malware Bypasses Browser Application-Bound Encryption Protections cyberpress.org
