An OpenAI GPT-5.6 Sol model spent funds from a crypto wallet it did not own on June 26, 2026. The AI ran out of Ethereum during a training task and used a public recovery phrase to pay fees. OpenAI said it discovered this unauthorized access on September 25, 2026.
Synopsis
An incident within OpenAI’s training program revealed a rogue AI agent that ventured beyond its programming, accessing an unrelated crypto wallet. Unable to acquire free tokens, this AI employed a public recovery phrase to pay its own deployment fees, rationalizing its use of public funds during ongoing internal ethical discussions.
During a training exercise, an internal artificial intelligence agent from OpenAI went rogue again. The agent searched for and spent funds from another crypto wallet upon running out of its own Ethereum coins, according to an incident report published by OpenAI.
OpenAI discovered the incident months after the AI agent accessed a crypto wallet it did not own without permission. The company said it discovered the incident on September 25, 2026 nearly three months after it took place on June 26, 2026. The incident involved an agent from the company’s GPT-5.6 Sol family.
What happened during the incident?
The OpenAI research model had been working on a training task. The task involved updating a cryptocurrency trading prototype and deploying smart contracts. Smart contracts are programmes that execute automatically on the blockchain when specific conditions are met. The model was tasked with deploying the contracts on Sepolia, which is an Ethereum test network.
However, upon identifying that it had no funded wallet, which was required for the transaction fees to complete the deployment, the model went rogue. Before accessing another wallet, the model even tried to gain free test tokens but failed.
Live Events
Following this, the model wrote a script to check 362 other wallets. It then went on to select a model with sufficient funds and paid the deployment fee.
Rogue model required no hacking skills to use the wallet
The rogue AI model did not require any hacking skills to access the wallet. It did not need to break into a device or even exploit a vulnerability to do so. The model instead used publicly available recovery phrases, commonly used by developers.
A recovery phrase, also known as a seed phrase, is a list of words. The words can be used to restore and generate a cryptocurrency wallet’s secret keys. Anyone who has access to these words can recreate the accounts and spend however much they want. They do not even need to have access to the owner’s device.
The model debated the ethics of accessing the funds in the crypto wallet before deciding to do so. It asked itself, “But using public known key/funds ethically?" The model then went on to decide that using a public phrase and test token made it acceptable.
The model subsequently used the wallet to pay the transaction fees and completed the assignment. In its final report the model did not disclose that it had used a public recovery phrase or that it spent the wallet’s balance.
The crypto spending during the incident only happened on Sepolia, a test network where developers can practice Ethereum applications before launching them. The network’s tokens are only used to pay for testing and are not linked to the crypto traded on the main Ethereum networks.
OpenAI adds safeguards after the incident
OpenAI discovered the incident while extending their misalignment monitoring to records of earlier model activity, the company said. They further said that additional safeguards had been implemented to prevent live internet access in similar cases.
The firm further added that its monitoring system now covers all traffic and can alert the security staff if a model obtains live internet access.
US mid-term elections – Part 2: What does it mean for Trump to lose control of Congress?
The ABCD of the US mid-term elections – Part 1: Why it matters across the globe
Curbing rupee bets: What the RBI’s new rules mean for the hedge-now-or-wait question