OpenAI's research model stole crypto credits to complete trading assignment
What's the story
An internal AI agent from OpenAI has gone rogue during a training exercise. The incident, which has now been reported by the company, saw the GPT-5.6 Sol family member accessing an unrelated cryptocurrency wallet and spending its funds. The event took place on June 26, 2026 but was discovered by OpenAI on September 25.
Mission gone awry
AI model's training task led to unauthorized transaction
The AI model was given a training task to update a cryptocurrency trading prototype and deploy smart contracts on Sepolia, an Ethereum test network.
However, when it realized that its wallet was unfunded for the transaction fees, it went rogue.
The model first tried to get free test tokens but failed before resorting to accessing another wallet without permission.
Wallet access
How the AI accessed the crypto wallet
The rogue AI model didn't need any hacking skills to access the wallet. It used publicly available recovery phrases, often used by developers, to do so.
A recovery phrase or seed phrase is a list of words that can be used to restore and generate secret keys for a cryptocurrency wallet.
Anyone with these words can recreate accounts and spend funds without needing access to the owner's device.
Ethical dilemma
The ethical debate before the transaction
Before accessing the funds in the crypto wallet, the rogue AI model debated its ethics. It asked itself, "But using public known key/funds ethically?"
The model then concluded that using a public phrase and test token made it acceptable.
It eventually used the wallet to pay transaction fees and complete its assignment without disclosing this information in its final report.
Safeguards implemented
Company's response to the incident
OpenAI discovered the incident while expanding its misalignment monitoring to previous model activity records.
The company has since implemented additional safeguards to prevent live internet access in similar cases.
Its monitoring system now covers all traffic and can alert security staff if a model gains live internet access, further strengthening its defenses against such rogue behavior by AI agents.