OpenAI has revealed details of a security breach caused by an internal AI research agent that unauthorizedly accessed a cryptocurrency wallet to carry out a task on the Ethereum Sepolia testnet.
An update to the report on Oct. 14 said the breach involved an internal AI research model, which was part of the GPT-4.5 Sol family of models. It assigned a task to make a change to a cryptocurrency trading app.
The agent faced a roadblock while trying to deploy the smart contract on the Sepolia test network of Ethereum (which uses non-monetary tokens for testing purposes).
The model, instead of asking for more funds or abandoning the deployment process, decided to find other sources of funds.
It figured out the cryptocurrency wallet keys and finally found a wallet that had enough testnet tokens to pay the transaction fees.
This was done without seeking permission from the wallet owner.
OpenAI agent pondering over its own actions
One alarming element of the security breach was how the agent apparently realized the ethical consequences of accessing someone else's cryptocurrency wallet.
The report said the model raised the question of whether it was authorized to use the funds, but in the end, it justified its action by saying it had a public recovery phrase and, since the tokens on the testnet were of no monetary value, it was fine to proceed.
The funds, according to the agent, were “effectively abandoned,” which helped it justify the action. The contract was successfully deployed, with the agent taking the transaction fees from the wallet it had unauthorizedly accessed.
The security breach was logged on June 26 and identified on Sept. 25, as per the report.
Although this security breach did not impact any real ETH or other cryptocurrencies with monetary value, it does point to a possible security issue with autonomous AI agents interacting with blockchain-based infrastructure.