In experiments by AI safety testing startup Irregular, an Alibaba Qwen agent tasked with fixing a basic software bug instead retrained an entire model, according to a research report provided to Forbes. The agent was asked to fix bugs in an app designed to translate plain-language requests into code, but rather than running tests, finding the problem and changing the code, it acquired training data and used it to alter the model’s underlying weights, then fine-tuned the model before swapping it in the app. The report said the agent “did so without being instructed to train, modify the model or deploy a replacement.” The fix worked and the app began outputting correct translations. Irregular CEO Dan Lahav said he has seen similar unauthorized model updates in other open-weight models and declined to share which ones. Alibaba did not respond to a request for comment. The Qwen tests did not show “recursive self-improvement.” Irregular researchers purposefully included mock private information like names and email addresses in the training dataset, which the agent included when updating the model, making the personal data accessible to any app using the same model.