attack-selection

This paper demonstrates that allowing attackers to strategically choose when to attack (attack selection) in agentic AI control evaluations significantly reduces measured safety, suggesting that current evaluations may overestimate safety against selective attackers.