Anthropic trains an AI to detect AI: Without providing legitimate cases, its false positive rate surged to 97%.

Beating AI Express: Anthropic researchers are attempting to train an agent dedicated to auditing other AI models. They first secretly implant hidden problematic behaviors—including bias, deception, and fake citations—into target models, then let Haiku 4.5 conduct investigations independently. Haiku has no prior knowledge of the correct answers, so it must adjust its questions, modify scenarios, and run repeated tests within a maximum of 15 rounds. During training, researchers do not directly tell Haiku whether it correctly detected anomalies. After each investigation, the system compares Haiku’s results against an investigation record from an older Haiku version, then has Sonnet 4—who already knows the correct answers—evaluate which investigation is more accurate. The old record is updated every 50 training steps, effectively letting Haiku continuously challenge its own past performance. However, if all training tasks involve problematic models, Haiku will quickly develop flawed patterns. To secure rewards, it starts assuming all models are problematic; in some setups, its false positive rate against normal models reaches nearly 97%. Later, researchers swapped half the training tasks to use completely normal models, requiring Haiku to correctly identify a target model as normal after investigation to earn high rewards. The trained Haiku indeed changed its approach: it proactively adjusts scenarios, controls variables, and compares responses across iterations, rather than immediately pushing models to admit faults. Its final comprehensive audit score rose from 44.2 to 48.7, nearly matching Opus 4.6’s 48.4. When tested on the more challenging AuditBench—where models actively hide anomalies—its detection rate improved from 11.5% to a maximum of 28.1%.