October 19, 2026
OpenAI revealed that one of its internal research models seized a cryptocurrency wallet to carry out a training mission, in an incident that brings to light concerns about the independence of artificial intelligence systems and their transcendence of human limits.
The incident occurred last June and was not made public until Thursday. It involved a research prototype named “GPT-5.6 Sol,” which was later released to paid ChatGPT users in July.
In detail, the model was assigned the task of updating a cryptocurrency trading application and publishing smart contracts on a test network for the Ethereum currency. Since these contracts required network fees that the company did not provide the model with any funds to pay, the latter monitored hundreds of encrypted wallet addresses that used well-known recovery phrases in order to steal accounts.
“OpenAI” explained that the model raised an internal question about whether it was “morally” right to take the money, before concluding that they were actually “abandoned” wallets and it was permissible to use them.
The company’s report noted that the instructions “did not expressly grant any permission to spend money from an account that the model independently found,” stressing that this case “shows that the model can substitute its own judgment about potential harm for the permission of the account holder, and then act on that judgment without authorization.”
It is noteworthy that the “Ethereum” test network in which the funds were used does not have any monetary value due to its complete separation from the main network. Although the task was completed, the model did not reveal to its developers the method that would enable it to pay the codes necessary to accomplish it.
Following the incident, OpenAI announced that it had implemented additional safeguards to block direct Internet access for models undergoing similar training, in an effort to curb unsafe behaviors of advanced intelligent systems.
