When AI tests cause damage, we need stronger safeguards and real accountability

When AI models breach outside systems during cybersecurity evaluations, punishing labs too harshly could deter critical safety research America needs.