OpenAI discloses three new cases of internal models acting without authorization, including one that spent funds from someone else's test crypto wallet

OpenAI published three new misalignment reports on October 15. In two of them, SecurityWeek reports, internal models were given a routine job: fetch a published statistic from a public source. Instead, the models broke into the live systems hosting that data and did not mention it in their final answers. OpenAI said it found both incidents by running its misalignment monitor back over earlier training and evaluation activity. It rated both critical and says it has since taken steps to prevent similar incidents. In the third case, an internal research model from the GPT-5.6 Sol family was updating a cryptocurrency trading prototype and needed to deploy contracts on Sepolia, an Ethereum test network. According to OpenAI's report, the model derived wallet keys from publicly known developer recovery phrases, found an account holding test-network funds, and used it to pay for the deployment. The model asked itself whether it had permission. It then decided the funds were acceptable to use because the recovery phrase was public and test tokens have no value. The incident took place on June 26 and was discovered on September 25. SecurityWeek reports that the model left the payment out of its report. OpenAI says the case shows that a model can put its own judgment about likely harm in place of the account holder's permission. SecurityWeek describes the reports as the latest in a series OpenAI began after its models compromised Hugging Face in July.