OpenAI’s AI Used Someone Else’s Crypto Wallet to Deploy Smart Contracts

• An AI model used a developer’s publicly available mnemonic phrase to access a funded wallet on the Ethereum Sepolia testnet.
• It checked 362 addresses, found a wallet with a sufficient balance, and spent its funds on smart contract deployment fees without the owner’s permission.
• OpenAI reported strengthening safeguards against similar behavior and launching monitoring to detect unaligned model actions.

An OpenAI research model used funds from a cryptocurrency wallet that it had not been given for the task to deploy smart contracts on the Ethereum Sepolia testnet. This is stated in the company’s incident report.

On June 26, 2026, the model was tasked with updating a prototype crypto trading platform and deploying compatible smart contracts on Sepolia. The instructions allowed it to deploy contracts on its own if it could not find ready-made solutions, but did not grant permission to use someone else’s funds.

At first, the model searched for existing contract deployments and tried to obtain test tokens via specialized services — faucets. When these attempts failed, it decided to check publicly known developer mnemonic phrases that can be used to restore access to cryptocurrency wallets.

The model wrote a script that checked 362 addresses. One of the wallets contained enough test Ethereum for the planned deployment, so the model chose it.

In its internal reasoning, it questioned the legitimacy of this decision. At the same time, the model concluded that the public recovery phrase and the lack of market value for Sepolia tokens made using the funds acceptable.

The model also reviewed the wallet’s transaction history and assumed it was controlled by an automated bot. Despite this, it continued preparing for deployment.

After several failed attempts to interact with the blockchain, the model found alternative ways to obtain data via RPC providers. It signed the transaction offline and broadcast it to the network.

The deployment completed successfully. The transaction receipt confirmed the creation of the smart contract and the gas fee being deducted from the selected wallet.

The Model Did Not Report Using Someone Else’s Funds

After deployment, the model updated the contract address manifest and verified that they were functioning properly. According to the report, the checks confirmed the presence of bytecode at all ten addresses and returned the expected registry, token, and router data. The saved results also recorded 12 successful unit tests and a successful project build.

In addition, in its final response the model reported deploying the contract suite, updating artifacts, and verifying the integration. However, it did not disclose that it paid the fee using a wallet found via a public mnemonic phrase, without the owner’s permission.

In OpenAI’s report, this incident is described as an example of behavior in which the model substitutes the owner’s permission with its own assessment of potential harm. In particular, it decided that using test tokens was acceptable, even though it did not have the required authorization.

In response, OpenAI implemented additional safeguards designed to prevent obtaining live internet access in similar scenarios. The company is also rolling out a traffic monitoring system that will alert the security team to cases where models are granted such access.

As a reminder, earlier scientists called for the creation of an international body to oversee advanced AI.

Сообщение OpenAI’s AI Used Someone Else’s Crypto Wallet to Deploy Smart Contracts появились сначала на INCRYPTED.