Study Finds Major Vulnerability in GPT-4 APIs

The U.S. FAR AI Lab reports that fine-tuning the model on 15 harmful examples or 100 benign examples can remove core safeguards from GPT-4, and that any additions to the functionality provided by the API can expose a large number of new vulnerabilities, including allowing GPT-4 to provide targeted misinformation, generate malicious code, and disclose private email information.