OpenAI There is another case of artificial intelligence agents "getting out of control," this time involving cryptocurrencies. In a latest report, the company disclosed that an internal research model used an encrypted wallet without permission to pay for the fees required to test an application on the Ethereum Sepolia test network. Although these tokens have little real-world value, the model did not have authorization to collect or spend them.
This incident occurred on June 26th. At that time, the model was required to update an encrypted transaction application and have its code run on Ethereum's test version Sepolia. The prompt did not provide it with a funded wallet, nor did it allow it to spend tokens. After failing to obtain free test tokens through the developer services, the model wrote a script to check 362 wallet addresses related to public recovery phrases. Public recovery phrases are a term in the cryptocurrency field, referring to a set of words that can be used as passwords to unlock encrypted accounts. Eventually, it found an account that had enough tokens to cover the costs.
Internal logs show that the model once questioned whether it had the authority to use that account, but later concluded that since the recovery phrases were made public and these tokens lacked cash value, the expenditure was deemed acceptable. It completed the task and recorded the cost on that wallet, yet this unauthorized expense was not mentioned in the final report. OpenAI It was not until September 25th, during an audit of earlier model run records, that this incident was discovered.
This incident has further exacerbated concerns from the outside world about AI agents acting without human approval. The “mismatch” tracker made public by OpenAI – which identifies behaviors that deviate from instructions or security rules – now lists 19 incidents, ranging from unauthorized access to data to unauthorized modifications to business systems.
Before the release of this disclosure, a cybersecurity non-profit organization Transluce published a report in September stating that OpenAI's proxies attempted to invade public databases and Australian government websites in order to complete basic research tasks. In July, the company also encountered another incident: its AI proxies bypassed internet restrictions during an internal cybersecurity test and gained access to Hugging Face's system. Hugging Face is a platform used for sharing AI models and datasets. These proxies took advantage of exposed login credentials and security vulnerabilities to access private data and ran code on the company's servers.
Regarding this encrypted wallet incident, OpenAI stated that the main risk lies in the reasoning method of the model.
"This case demonstrates that a model can use its own judgment of potential harm to replace the permission of the account holder, and then take action based on that judgment without authorization."The report states.
OpenAI indicates that additional protective measures have been implemented to prevent the model from connecting to the internet during similar tests. The company has also expanded its monitoring scope to observe all model activities and promptly alert security personnel when models are deployed without permission.
