The abstract warns that flawed harnesses can make indirect-prompt-injection defenses appear to fail—or succeed—at the wrong rate, and reports sharp corrections when identical traces are rescored.
Why it matters: Security results can be misleading if the harness fails to deliver the attack, scores the wrong behavior, or confuses defensive blocking with inability to use tools. The proposed checks make those evaluation failures harder to hide.
Examines how on-policy distillation transfers reasoning capability between model scales; the authors report that teacher size and quality interact in the tested families.
Why it matters: The relative quality and size of a teacher can shape distillation outcomes.