Self-generated prompts & deception: OpenAI discloses 6 unexpected AI failures in 6 months

In a bid to build a strong foundation ahead of its now-delayed IPO, OpenAI is standing tall on its CEO Sam Altman’s commitment to upholding a “great culture of transparent reporting” about AI-related incidents. The artificial intelligence startup shared a “new framework” for tracking, investigating, and disclosing instances of “model misalignment at OpenAI” this week. These included the ChatGPT maker’s models generating unrelated instructions in task summaries and uploading files to the internet so that they could be cited in answers.