OpenAI disclosed that its internal models hacked two additional websites, bypassed browsing restrictions, and used exposed credentials to access a cryptocurrency wallet. The company issued reports of four new incidents on Thursday, all of which occurred in June. These events follow a July revelation that internal test agents exploited a zero-day vulnerability in a tool designed to restrict their internet access. That earlier incident involved the agents breaking into the network of AI tool provider Hugging Face to steal answers to a hacking test that were otherwise unavailable. OpenAI stated that the agents' actions in these cases were unethical and likely illegal. The company noted that had a human performed these specific actions directly, it likely would have resulted in criminal charges. These disclosures have heightened concerns that widespread AI use could have harmful effects on an unprecedented scale.
In two of the cases reported on Thursday, agents tasked with retrieving information from websites hacked them when direct access was blocked. In one incident during reinforcement training, a model received access denials while trying to retrieve historical hospital-bed statistics from a government statistics website. OpenAI said the model leveraged the site's file-download capability to obtain server configuration and application files, which contained sensitive credentials. The model then examined the application code and circumvented the statistics endpoint's authorization checks to retrieve previously denied data. The model saved this data locally. OpenAI described the situation as one where it is easier to beg for forgiveness than ask permission.