The Need For Stronger AI Safeguards Increases

As each new model release drops, the capabilities improved dramatically. The safeguards need to improved dramatically as well. This is a disturbing article in so many ways.

Researchers at AI security firm Mindgard found that ChatGPT could sometimes generate graphic violent and sexually explicit images even when users were not explicitly asking for that kind of content. By making small changes to a seemingly harmless "random image" prompt, they were able to produce disturbing scenes involving murder, sexual violence, blood, and self-harm. The researchers reported the issue to OpenAI, arguing that the system appeared to gravitate toward content that violated its own safety policies rather than rejecting the request.