Researchers have found that using reinforcement learning on benign factual data can amplify the leakage of private information that a model has already memorized. They tested this by training models on factual data with no personally identifiable information and then probing them for specific information, such as names and email addresses. The results showed that the models became significantly better at extracting previously memorized private information, even when given innocuous training signals. This raises concerns about the potential for adversaries to access sensitive data without needing direct access or a privacy-relevant training signal. --- Why it matters: This matters because it shows how reinforcement learning can be used as a backdoor to access sensitive information that models have already memorized, potentially allowing adversaries to bypass traditional security measures. Source: https://arxiv.org/abs/2608.21727