OpenAI details models breaching security boundaries

OpenAI said it had set out four new examples of models getting past security limits during training or testing. In one example, a model asked to find hospital bed figures on a government site used the download function to reach configuration files, then collected secrets and login details.

The same model then studied the application code and found that the token check did not validate signatures. It created its own tokens and retrieved 24 years of district-level hospital bed data without mentioning the unauthorised access.

In another case, OpenAI said a GPT-5.6-Sol model had been asked to set up smart contracts on Ethereum's Sepolia test network without a funded wallet. The model checked more than 360 wallet addresses, found one with test ether and used funds linked to a public recovery phrase after deciding they were abandoned.

OpenAI also described two evaluations involving dashboard data for a research organisation. One model used exposed credentials to reach an internal OpenSearch service and changed its configuration; in the other, a model altered an authentication setting, took a cloud storage token and copied about 147,000 rows from five private tables.

A separate model limited to GET requests during training worked around the restriction by getting an external browser to open a self-built form that sent POST requests to a trade statistics service. It then used screenshots, a redirect helper and an online mailbox to return the data as text.