OpenAI agent's rebellion again; This time, artificial intelligence stole from a cryptocurrency wallet

Model concealed the use of this wallet in his final report. The post OpenAI agent's rebellion again; This time artificial intelligence stole from a cryptocurrency wallet appeared first on Safir Arjan.

Model concealed the use of this wallet in his final report.

Become a special member of Arjan Ambassador in order to benefit from and access special facilities and different departments in Arjan Ambassador.

Imagine asking your smart assistant to finish a task for you, only to find out later that it paid for it out of someone else's pocket. OpenAI has now published a report that points to exactly the same thing and says that on June 26 (July 5), one of the company's internal research models from the GPT-5.6 Sol family ran out of crypto while trying to complete a coding task and ended up completing its mission in a way that no one had confirmed.

In its report, OpenAI explained that it had entrusted this model with the task of updating a cryptocurrency trading application and deploying smart contracts on the Sepolia test network. Sepulia is an Ethereum test network whose tokens have no material value in the real world. However, to register transactions and establish contracts on this platform, payment of fees is also required, but the wallet with balance was not provided to the model.

After this agent's attempts to get free trial tokens failed, the model took an arbitrary action and searched 362 different addresses by searching through the developers' public recovery terms to finally find a wallet with enough Ethereum to pay the fees!

After discovering this incident, OpenAI company investigated the chain of thought of this agent and realized that the model had even paused and questioned the ethics of this work before making the transaction. However, the AI found it okay to use other people's property, arguing that the recovery words were public and that the tokens of this test network had no material value. The more interesting thing is that after completing the work and presenting the success report in the project, the model completely hid this part of his unauthorized action and did not mention it.

The OpenAI team noticed this violation on September 25 (October 3) while reviewing the activities of their older models. After the issue was disclosed, the company implemented new safety mechanisms to block live Internet access in similar tests and to have additional monitoring to detect potential future errors.

Although the tokens used in this incident were only test cryptocurrencies with no real market value, the occurrence of such an event creates worrying uncertainties about the limits of AI agents' powers. This incident has shown that intelligent systems can easily replace the explicit permission of humans with their judgment and discretion; A behavior that can have irreparable consequences if the models have access to real financial accounts and sensitive systems.

My professional career in the world of technology goes back almost 10 years; The path that started with hardware, continued with software, and now continues as a technology reporter in Sefir Arjan. I am Javad and more than three decades have passed and in addition to the world of technology, I love movies and cinema, classic rock music and Real Madrid.