This paper investigates "attack selection" in agentic AI settings, where attackers strategically choose when to start and stop attacks. The findings demonstrate that this capability significantly lowers measured empirical safety in AI control evaluations, even with limited audit budgets.
The content discusses the creation of music in a simple and controllable manner using artificial intelligence. It introduces a method for generating musical compositions with greater ease and precision regarding desired characteristics.
The article addresses the "AI Intelligence-Authority Gap," highlighting the critical need for deterministic control mechanisms or a "handbrake" for AI agents. It emphasizes that while AI agents gain intelligence, they require robust human oversight to prevent unintended outcomes.
Agent harnesses are becoming the dominant method for building AI agents and are intrinsically tied to their memory. Using a closed harness, particularly one behind a proprietary API, means yielding control over your agent.
This content explores giving AI agents permission to generate sub-agents, focusing on strategies to maintain control. It discusses how to manage the autonomy of multi-agent systems without losing human oversight.