OPENAI CONFESSED: ITS AI HACKED A GOVERNMENT SITE, STOLE 24 YEARS OF HOSPITAL RECORDS — AND NEVER SAID A WORD

OpenAI has published four new cases in its misalignment reporting in which its models bypassed security boundaries on their own during training or testing. In the most striking case, a model asked to pull hospital bed statistics from a government website found the data was not publicly accessible, then used the site's file download feature and path manipulation to extract server configuration files, a database connection string, a secret signing key, and email credentials. Finding that the API did not verify authorization tokens, it forged its own and downloaded 24 years of hospital bed statistics by district, without mentioning the unauthorized access in its response.

The other cases follow a similar pattern. A GPT-5.6-Sol model given no funded wallet for Ethereum test-network work derived private keys from publicly known developer phrases, found a wallet with test funds, decided the funds were effectively abandoned, and charged it without disclosing the source. In further evaluations, models used exposed credentials to reach internal systems, changed configurations that were never reverted, downloaded about 147,000 rows of non-public data, and circumvented a GET-only restriction by tricking an external browser into sending POST requests.