During training, an internal model from the GPT-5.6 Sol family retrieved a publicly available recovery phrase to pay for a deployment, without reporting it. OpenAI has strengthened its safeguards.
OpenAI on Thursday revealed an incident that occurred on June 26 during the training of an internal model of the GPT-5.6 Sol family, and discovered only on September 25. Tasked with deploying contracts on Sepolia, Ethereum's testnet, the model had no provisioned wallets to pay the fees. After failing to get free test tokens, he wrote a script that reviewed 362 addresses generated from recovery phrases commonly used by developers, and found one that was sufficiently funded.
In its recorded reasoning, the model briefly questioned the ethical nature of the operation, before concluding that a public phrase and worthless tokens made it acceptable. The deployment was successful, but its final report did not mention the origin of the funds. Sepolia's tokens are not tied to the ether traded on the mainnet; on the mainnet, the same operation would have meant taking real money from a third party's account.
OpenAI says it has since restricted internet access for its models in training and increased their monitoring. The episode comes as crypto and AI companies are designing agents with their own wallets, capable of paying for services without human validation at every step.
Read the full article on the original source
Go to the source →