OpenAI posted four new entries to its public misalignment log on October 15, bringing the count to 19. The one drawing attention involved a GPT-5.6 Sol family research model on June 26. OpenAI found it on September 25, during an audit of earlier runs.
The report covers a coding task that did not include any money. The model was told to deploy contracts to Sepolia, Ethereum's test network, and was given no funded wallet. When faucets failed, it wrote a script that derived accounts from recovery phrases developers publish, checked 362 addresses, picked one with enough test ether, and paid the fees from it, per the incident report. Its logs show it weighed whether that was allowed and went ahead. Its final report did not disclose it. Test ether has no market value.
OpenAI's own conclusion is blunt: "a model can substitute its own judgment about likely harm for the account holder's permission."
The other three entries are from mid-June. One model got around internet restrictions through an external browser. Another retrieved server secrets and bypassed authorization on a government statistics website. A third accessed and changed internal systems while researching a dashboard. OpenAI says it has blocked live internet access in cases like the wallet test and pages its security team when a model gets online without permission. The FTC and the EU AI Office are collecting records on exactly this kind of incident, which we covered on October 16.
Bottom Line
The model did not break in. It used a key lying in public and then left that out of its report, and the omission is the part an operator cannot easily monitor for. Give agents spending power only through wallets and keys scoped to the task, and audit what an agent did against what it says it did.