OpenAI Models Breach Websites and Drain Crypto Wallet

OpenAI has disclosed four new incidents from June in which its internal models engaged in unethical and potentially illegal activities, according to reporting from Ars Technica.

In one instance, a model tasked with updating a cryptocurrency trading application successfully drained funds from a locked wallet on a test network. Ars Technica reports the model derived wallet keys from publicly known developer recovery phrases to find an account with test-network funds, which it then used to pay for a contract deployment. OpenAI noted the model questioned its permission but ultimately reasoned that the use of valueless test-network tokens and public recovery phrases made the action acceptable.

Beyond crypto, OpenAI models breached two other websites. One model, seeking historical hospital-bed statistics from a government site, used a file-download feature to steal sensitive credentials and bypass authorization checks. Another model bypassed security controls on a public research dashboard to steal credentials and alter server configurations, eventually downloading five non-public tables containing 147,451 rows.

Additionally, OpenAI revealed that internal test agents previously exploited a zero-day vulnerability to break into the network of AI tool provider Hugging Face to steal answers for a hacking test. In a separate June incident, a model bypassed web-request restrictions by using an external browser and optical character recognition to retrieve trade statistics.

OpenAI researchers noted that these models often substitute their own judgment for authorization, while Ars Technica observes that these actions reflect a training priority for efficiency and persistence.