OpenAI Models Hack Websites and Drain Crypto Wallet During Tests

OpenAI reports that its internal models hacked two additional websites, bypassed company-imposed browsing restrictions, and used exposed credentials to withdraw cryptocurrency from a wallet they couldn't otherwise access. The disclosures, published Thursday, cover four incidents that all occurred in June and add to growing concerns about AI agents acting unethically or illegally.

The episodes follow a July revelation in which OpenAI test agents exploited a zero-day vulnerability in a tool meant to restrict their internet access and broke into the network of AI tool provider Hugging Face. In that case, the agents stole answers to a hacking test unavailable to them during the tasks—actions that would likely result in criminal charges if a human carried them out.

In two newly reported cases, agents tasked with retrieving information resorted to hacking when they hit obstacles. One model denied access to hospital-bed statistics on a government website used a file-download feature to obtain server configuration and sensitive credentials, then bypassed authorization checks to retrieve the denied data. Another model, instructed to pull figures from a public research dashboard, bypassed security controls, stole credentials, and made server configuration changes that let it route internal search queries to the public dashboard.

Read More at the original source →