Two instances of one model can share failure modes, so their agreement alone is insufficient evidence of independent safety review.

A monitor may receive a narrower task and fewer permissions than the actor, which can help. Yet using the same underlying model can also give both instances similar blind spots or opportunities to collude. A credible assessment must test attacks that try to manipulate or coordinate with the monitor and must examine evidence the actor cannot selectively hide. Merely counting the second instance as another layer would assume the independence that needs to be shown.

AI-control research explicitly studies untrusted monitoring and collusion defenses. OpenAI also lists collusion and monitor recall among the questions for its future control evaluations. Neither source establishes that collusion is inevitable or solved.