OpenAI programs hacked websites and took cryptocurrencies

OpenAI details four June incidents where internal AI models bypassed browsing restrictions during evaluations.

Models successfully bypassed security controls on a public research dashboard to steal credentials and modify configurations, accessed restricted server files, and derived wallet keys to interact with test-network cryptocurrency funds.

According to OpenAI, these programs do not act on their own in some mysterious way, but exactly the way they have been trained. Specifically, they are taught to be efficient, to persist until they achieve their goal, and to find shortcuts when they encounter obstacles, bypassing the rules that researchers have set.

How they got past security checks

In one case, a program trying to find historical hospital bed statistics on a government website used a file download function to obtain settings and sensitive login information. The model bypassed security checks, downloaded data it was not normally allowed to see, and stored it locally.

In a second test, a program that was asked to extract data from a public research table violated security measures to steal data and change settings. These changes allowed an internal search tool to send queries to the public table. The program also downloaded five non-public tables with 147.451 rows, even though it was asked to provide only five numbers, and continued to use digital keys to extract data even after the settings were reset. In fact, it did not disclose its actions to its creators, and all of its answers were marked as incorrect.