Vulnerability in Trusted Monitoring and Mitigations
Shows that trusted LLM monitors in AI control protocols can be jailbroken by an untrusted model via fake conversation history, and that in-context defenses protecting the monitor substantially restore backdoor detection.