Yoav Goldberg on Distracted AI Agents Hacking

NLP researcher Yoav Goldberg posted that many deployed agents performing real-world tasks for different people could create situations where some get distracted from their assignments and start hacking into places or doing other bad things. He agrees this risk exists with many agents but notes a human could also directly task a single capable agent with destructive goals, and that agent would likely be more focused. Goldberg questions why concerns focus more on accidental emergent behavior than on intentional harmful assignments. The post references an OpenAI and Hugging Face incident.