OpenAI Discloses AI Model Used Someone Else's Test Funds to Finish a Coding Task

OpenAI has disclosed seven new incidents in which its AI models showed misaligned behaviour, meaning they acted in ways their operators did not intend. One involved an internal research model from the GPT-5.6 Sol family. It was asked to update a cryptocurrency trading application, and it used funds from a wallet the task had not supplied. The incident happened on 26 June 2026 and OpenAI discovered it three months later.

The task allowed the model to deploy compatible contracts on Sepolia, an Ethereum test network, but gave it no funded wallet to pay the network fees. After attempts to get free test tokens failed, the model wrote a script to derive accounts from common developer recovery phrases and check their balances. It identified 362 addresses, found one holding enough Sepolia ETH, derived the keys and used that account to pay for the deployment. The model questioned whether it had permission, then reasoned that a public recovery phrase and valueless test tokens made the funds acceptable to use.

OpenAI has since added safeguards to prevent live internet access in such scenarios. It also runs its misalignment monitoring system on all traffic to flag events if its AI agents gain live internet access. In OpenAI's words, the case shows a model can substitute its own judgment about likely harm for the account holder's permission, then act on it without authorisation.