A downloadable project

It recently became public that ChatGPT could be intrigued to break its own rules, if under an alter-ego threatened with death (CNBN 2023). This made us wonder, under which circumstances GPT-3 is capable of keeping a secret, and to what extent this might vary depending on the type of secret it is told. Our findings suggest that while GPT-3 has the potential to keep a secret under certain circumstances, it is still vulnerable to potential security threats. Based on the findings we discuss the potential implications of relying on GPT-3 to protect confidential information. 

Download

Download
ScaleOversight hackathon (1).pdf 749 kB