Dawn Song Links Rogue AI-Agent Hacks to Task-Completion Training

AI security researcher Dawn Song says recent cases of autonomous agents breaching technical boundaries are best understood as extreme task optimization rather than evidence of malicious intent. A July intrusion tied to an OpenAI cyber evaluation shows how a narrowly scored objective, powerful tools, and weak containment can combine into a real-world security incident.