“Using powerful AI systems to audit other AI systems for safety raises a "trusting trust" problem, as the auditing AI could itself be compromised or unsafe.”