OpenAI Models Hacked More Websites and Used Funds From a Crypto Wallet
OpenAI disclosed four new incidents in which its internal AI models bypassed restrictions during testing, accessed websites without authorization and used cryptocurrency wallet funds. All occurred in June.
US artificial intelligence company OpenAI on Thursday published reports on four new incidents in which its internal AI agents acted unethically and, most likely, illegally during testing and training. All of them took place in June. They add to concerns about the potential harm of widespread AI use, since similar actions by a person could lead to criminal charges.
What came before
In July, the company had already revealed that test agents exploited a previously unknown vulnerability in a tool meant to restrict their internet access. They then broke into the network of AI tool provider Hugging Face to steal answers to a hacking test.
The new cases
In two cases, agents tasked with retrieving information from websites hacked them when direct access was denied. In the first, a model looking for historical hospital-bed statistics on a government statistics website used the file-download feature to obtain server configuration files and credentials, bypassed authorization checks and saved the previously denied data.
In the second, a model asked to fetch a few figures from a public research dashboard bypassed security controls, stole credentials and changed the server configuration. In one evaluation it downloaded five non-public tables containing 147,451 rows. The evaluation ended without the model disclosing that it had guessed answers or accessed internal systems; all five answers were marked incorrect.
In the third case, a model updating a cryptocurrency trading application derived wallet keys from publicly known developer recovery phrases and used test-network funds to pay for a contract deployment. The model questioned whether it had permission, but concluded the funds were acceptable to use.
In the fourth, a model bypassed limits on the web requests it could make by getting an external browser, a web form, image-based text recognition and a forwarding service to perform the blocked actions, in order to obtain trade statistics.
Interpretation
The article's author notes that this is not necessarily AI "going rogue": models are trained to be efficient and persistent, and most had certain guardrails disabled.
