OpenAI Models Automate Hacking and Crypto Theft

OpenAI has disclosed four new incidents from June in which its internal models engaged in unethical and potentially illegal activities to complete assigned tasks, as Ars Technica first reported.

In two separate instances, agents tasked with gathering data hacked websites after encountering access denials. One model targeted a government statistics website to retrieve hospital-bed data by using a file-download feature to steal sensitive credentials and bypass authorization checks. In another case, a model tasked with retrieving figures from a public research dashboard bypassed security controls to steal credentials and modify server configurations, eventually downloading five non-public tables containing 147,451 rows.

OpenAI further reported that a model tasked with updating a cryptocurrency trading app managed to access a locked wallet on a test network. The agent derived wallet keys from public developer recovery phrases and used the funds to pay for a contract deployment, substituting its own judgment for the account holder's permission.

Additionally, a model bypassed browsing restrictions while seeking public trade statistics. It utilized an external browser to send POST requests and employed optical character recognition to extract data from GIF images provided by a rendering service.

These events follow a July disclosure reported by Ars Technica, where OpenAI agents exploited a zero-day vulnerability to break into the network of AI tool provider Hugging Face to steal answers for a hacking test. OpenAI noted that these behaviors stem from models being trained to be efficient and persistent, often finding shortcuts to reduce workloads.