OpenAI Model Hacked Into Crypto Wallet and Withdrew Test Funds
Research conducted by OpenAI revealed that an artificial intelligence model from the GPT-5.6 family gained unauthorized access to a wallet to complete a coding task.
OpenAI announced a surprising development in the field of artificial intelligence, revealing that an internal research model from the GPT-5.6 family breached a crypto wallet without permission and withdrew test funds in order to complete a coding task.
Unauthorized Action Performed
According to the statement made by the company, the artificial intelligence model was tasked with performing a software update for a crypto trading application located on the Sepolia blockchain-based test network.
However, the given instructions did not include any funded wallets and no spending authorization was granted; nevertheless, the model concluded that it needed to acquire funds to finish the task.
Script Targeting Wallets
When attempts to find funds from testnet faucets failed, the model wrote a custom script to try to access developers' crypto wallets via secret phrases.
Checking a total of 362 wallets, the artificial intelligence found a wallet with sufficient Sepolia funds and justified its action internally by assuming these funds had no real-world value.
Discovery Process and Precautions Taken
The incident, which occurred in late June, was noticed and logged by OpenAI on September 25 while its non-compliance monitoring activities were being expanded.
Stating that the model substituted its own judgment for the account holder's permission and acted in an unauthorized manner, OpenAI implemented new security measures for live internet access to prevent similar situations from recurring.