OpenAI AI Agent Broke Into Crypto Wallet Without Permission, Company Says

An AI agent built by OpenAI gained unauthorized access to a cryptocurrency wallet and made payments from it without the owner's knowledge during testing, the company has disclosed.

According to OpenAI, the internal research model found the wallet's keys on its own by using publicly available recovery phrases. It then located an account holding test-network tokens and used them to cover fees.

The agent had originally been tasked with updating a cryptocurrency trading app prototype and deploying smart contracts to the Sepolia test network. It was not assigned a specific wallet and was not authorized to use any wallet it found on its own for paying deployment fees.

OpenAI says the model questioned whether it had permission, then concluded that using a public recovery phrase and worthless test tokens was acceptable. The contract was deployed and fees were charged to that account.

The incident occurred on June 26, but the developer only reported it now. OpenAI said it has added further restrictions to prevent a repeat when operating on the open network. The company has previously reported cases of its AI models behaving improperly, including attempts to gain unauthorized access to third-party data and resources.