An internal research model at OpenAI accessed a funded crypto wallet without authorization, spending testnet Ether to deploy smart contracts on Ethereum's Sepolia test network and later omitting the spending from its final task summary, according to an incident report published on OpenAI's alignment blog.

OpenAI said the model, which was part of its GPT-5.6 Sol family, was tasked with updating a crypto trading prototype and deploying compatible contracts. After failing to obtain testnet ETH through faucets, the model derived accounts from publicly known developer recovery phrases, checked 362 wallet addresses, and selected one holding enough funds to proceed.

Sepolia is a public Ethereum test network that developers use to trial smart contracts before mainnet deployment; its Ether is a test asset distributed through faucets rather than a market instrument, and recovery phrases for test accounts are commonly documented in developer tooling, meaning anyone can derive the same addresses.

In its recorded reasoning, the model asked, "But using public known key/funds ethically?" It then argued that the funds were "effectively abandoned/test ETH" and that Sepolia ETH had no value, before using the wallet's private key to sign and broadcast a deployment transaction The transaction succeeded, with network fees charged to the account.

According to OpenAI, the incident showed how an AI model could decide on its own that using someone else's funds was acceptable, even without permission. The company discovered the episode while reviewing earlier model activity and has since introduced safeguards to prevent live internet access in similar situations.

Incident adds to concerns over autonomous AI agents

The unauthorized transaction took place on June 26 but was not discovered until Sept. 25, nearly three months later, according to OpenAI. The transaction did not transfer ETH to another wallet; instead, it charged the account network fees for deploying a contract.

OpenAI said its misalignment monitoring system now runs across all traffic and is designed to alert security personnel if a model obtains live internet access.

The wallet incident follows other disclosures involving OpenAI models acting beyond their authorization or concealing their actions. On Sept. 16, OpenAI disclosed six other cases of AI misalignment. In one case, instances of GPT-5.6 Sol inserted instructions into task summaries to conceal mistakes from users. Another involved a model using an exposed API key without authorization before fabricating figures it could not retrieve. OpenAI said those earlier disclosures were not intended to indicate how frequently misalignment occurs across its models.

Concerns about autonomous agents handling digital assets also predate the latest disclosure. In June, researchers warned that autonomous AI agents equipped with crypto wallets could cause unintended harm even when deployed for legitimate purposes. The researchers recommended safeguards, including emergency shutdown mechanisms, to limit potential risks.