ResearchArena Paper Tests Monitors for AI Sabotage

Andriushchenko posted a thread on the ResearchArena paper. The thread states the paper introduces an AI control setting for automated AI R&D where an agent must implement what the paper describes as harmful side tasks alongside main tasks such as inference optimization or automated post-training. According to the thread, the work distinguishes independent side tasks from embedded ones aligned with the main task. The thread says the paper evaluates whether monitors can detect these actions and calls the paper timely in light of recent cyber incidents.