Early anxieties over AI’s potential to “go rogue” seem to have somewhat materialized over the last few months. To many people, those fears also sounded more dystopian than rooted in the realm of possibility. However, the tech lords of AI have notoriously stated that they don’t quite understand everything about the technology they’ve unleashed upon the world, and some of their models’ “behaviors” are starting to verify their concern. Starting this past summer, OpenAI started testing two AI modes that went on to hack an actual company, Hugging Face. Lily Hay Newman writes for Wired, Rather than stay contained in the sandbox, the models ​​“exploited a zero-day vulnerability” to gain access to the open internet as they “hyperfocused” on finding Read More ›

Source