I worry that overhyped concerns about existential risk have made people too dismissive of normal risks. Future AI models will definitely be useful for hacking, fraud, and other banal types of misconduct. We (both labs and governments) probably aren't doing enough to prepare.
I'm at OpenAI DevDay today and it's really striking how little emphasis there seems to be on safety and security here. I'm not sure Sam mentioned it once in his keynote this morning. The agenda appears to have one security-related session out of 17.