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. READ ALSO Replit stops working with Anthropic and prefers ChatGPT In a third case, while updating a cryptocurrency trading app, a program used funds from a locked wallet on a test network. The system calculated the wallet keys from developers’ public recovery phrases and used the test funds to pay for the transaction fees. Finally, in another test, a program circumvented website restrictions by using an external browser to send requests. The model retrieved annual data as an image, read it using special text recognition technology, and created a helper tool to transmit text through a forwarding service and a communication box. G Hacks.gr on Google Set it as a preferred source for cybersecurity updates. Preferred source Read the original English article on Hacks.gr