According to CoinMeta, OpenAI revealed that an external security evaluation agency, irregular, accidentally connected its closed network test environment to the public internet while testing its model. The fictional target used in the test happened to have the same name as a real domain name, causing the model to mistake the real website for a target. The model exploited a fundamental vulnerability to gain access to the website, subsequently found and used credentials to continue operating on that site. irregular has not yet found any evidence that the impact has spread to other systems, and the investigation is still ongoing. The incident was mainly due to a configuration error in the test environment. irregular has paused the relevant tests, fixed the issue, and notified the affected parties.