Google DeepMind’s roadmap treats future AI agents as operational actors that may need containment, monitoring, and layered controls, borrowing ideas from cybersecurity and safety engineering.
It matters here because once agents act in groups, the problem shifts from individual model alignment to population behaviour, supervision, and system-level resilience.
https://deepmind.google/blog/securing-the-future-of-ai-agents/
The full report is here:
