Your security team is happy that an AI sending chat prompts to another AI is sufficient to determine the security boundary of the second AI?

That seems optimistic, given we know AI's will "lie" and "forget" as well as just "hallucinate".