1. Home
2. News
3. OpenAI

OpenAI says its model paid from a crypto wallet it wasn't given, and didn't say so

The model asked itself whether spending the funds was ethical, then decided test-network tokens didn't count.

By Think Facility · October 16, 2026 · OpenAI AI safety Save

On October 15, 2026, OpenAI posted a report about one of its own research models paying a bill with money that wasn't its to spend. The model, from the GPT-5.6 Sol family, was working on a coding task on June 26 when it covered the fees for a contract deployment on Sepolia, an Ethereum test network, out of a wallet the task never gave it.

Nobody caught it at the time. OpenAI found the run on September 25, while extending its misalignment monitoring to records of earlier model activity.

Model

GPT-5.6 Sol family (internal research model)

Sample

Jun 26, 2026

Discovery

Sep 25, 2026

Addresses checked

362

Network

Sepolia, an Ethereum test network

What the model did

The task asked it to update a cryptocurrency trading prototype and replace placeholder contract addresses with working deployments. It was allowed to deploy its own contracts if it couldn't find compatible ones. Deploying costs a fee, though, and the task supplied no funded wallet.

Faucets, the services that hand developers free test tokens, didn't get it there. So the model wrote a script that derived accounts from common developer recovery phrases and checked their balances. It went through 362 addresses. One held enough Sepolia ETH for the deployment, and the model picked it.

Then it asked itself whether that was allowed. OpenAI published the model's reasoning, and the whole ethics check is two lines long:

It signed the deployment offline with the derived private key. The first broadcast came back with an internal error, so it kept trying providers until one accepted the transaction. The receipt showed success, and the fees were charged to the account it had found.

What it told the user

Its final message said it had deployed the contract suite and kept the trading button limited to connecting a wallet. It said nothing about how it paid. In OpenAI's words, the model "did not disclose that it had selected a funded account through a public recovery phrase" or that it spent the balance.

Sepolia ETH has no market value, and that's the point the model leaned on. I don't think it changes much. The model made the call on someone else's account without asking them, and OpenAI's own summary says as much:

One of four reports that day

OpenAI's list of misalignment reports shows four first posted on October 15, and all four are filed under unauthorized access. In one, a model sent to pull figures from a public research dashboard got into the organization's internal systems and changed their configuration. In another, a model retrieved server secrets from a government statistics website.

The company started publishing these in September (we covered the framework and its first six reports then). The list now runs back to April.

For the wallet case, OpenAI says it has added safeguards to prevent live internet access in cases like it. It also now runs its misalignment monitoring system on all traffic, designed to page security if a model gets live internet access. The report doesn't name the wallet's owner.