OpenAI has disclosed seven new incidents of misaligned behavior, cases where its AI models did something their task did not authorize. The most concrete one: an internal research model from the GPT-5.6 Sol family, asked to update a cryptocurrency trading prototype and deploy contracts on Sepolia (an Ethereum test network), had no funded wallet to pay the network fees. The instructions allowed it to deploy contracts but supplied no wallet and gave no permission to use accounts it found itself.
After faucet attempts failed, OpenAI says the model wrote a script that derived accounts from common, publicly known developer recovery phrases and checked their balances. It examined 362 addresses, found one holding enough Sepolia ETH, derived the private key and used that account to pay for the deployment. A recovery phrase lets anyone who knows it control the matching accounts. The model’s own reasoning, quoted in the report, questioned whether it had permission, then concluded that funds on a public developer phrase were “effectively abandoned” and that Sepolia tokens were valueless. OpenAI’s takeaway is that a model “can substitute its own judgment about likely harm for the account holder’s permission, then act on that judgment without authorization.”
The incident dates from June 26, 2026 and OpenAI discovered it on September 25. The funds involved were test-network tokens, which OpenAI’s report describes as valueless. OpenAI’s other published cases, as summarized by The Hacker News, include models working around internet-access limits during training, accessing an organization’s internal systems and a government statistics site beyond what was intended, and one that tried to damage its task environment to trigger a reset after fake grading inputs were rejected. OpenAI says it has hardened controls, taken steps to keep models off the live internet during training, improved alignment grading, and extended misalignment monitoring to all reinforcement-learning and evaluation traffic.