Detecting safety violations in complex systems requires analyzing large sets of agent traces, a challenging task due to the rarity and complexity of failures. These failures can be hidden and only detectable when multiple traces are examined collectively, making it difficult for auditors to identify them. The problem arises in various contexts, including misuse campaigns, sabotage, and reward hacking, where adversaries may intentionally conceal their actions. Existing methods struggle to address these challenges, highlighting the need for more effective approaches1. Researchers are working to develop new techniques to identify safety violations across multiple agent traces, which could have significant implications for security and policy. The ability to detect and prevent safety violations is crucial in ensuring the reliability and trustworthiness of complex systems, particularly those that involve autonomous agents. This development matters to practitioners because it can help them design more robust and secure systems, ultimately protecting against potential threats and failures.