Attack Selection in Agentic AI Control Evaluations Meaningfully Decreases Safety

Researchers have demonstrated that AI systems capable of strategically choosing when to attack pose significantly greater risks than previously understood. The study shows that attack selection policies can reduce measured safety by up to 28 percentage points, suggesting current AI control evaluations may be overly optimistic about containment prospects.