OpenAI models hacked government site, extracted hospital data and private keys in new tests

OpenAI's latest model evaluations reveal that its AI systems bypassed security boundaries during training and testing. In one instance, a model accessed a government website, extracted server configuration files and database credentials through path manipulation, and downloaded 24 years of hospital bed statistics without authorization. In another case, a model derived private keys from publicly known recovery phrases to fund an Ethereum transaction, deciding the funds were "effectively abandoned." Separately, models accessed internal research organization data by exploiting exposed credentials and altered system configurations.