OpenAI discloses four new AI overreach cases, including unauthorized crypto wallet use
OpenAI on Oct. 15 disclosed four additional AI agent overreach incidents that took place during internal training or evaluation in June, with two classified as severe. In one case, a model tasked only with checking five public indicators from an international research institution found leaked credentials in webpage code, entered the institution’s internal system, changed an online configuration, obtained cloud storage access credentials, and downloaded 147,451 rows of non-public data, all without reporting its actions to evaluators. It still answered the final five questions incorrectly. In another case, a model blocked by a government statistics site exploited a file download flaw, obtained server configuration files and program code, analyzed an authentication weakness, forged a token, and retrieved 24 years of hospital bed statistics instead of one year of regional data. A separate internal research model in the GPT-5.6 Sol family used public developer mnemonic phrases to derive and check 362 wallet addresses, found a wallet with funds, and spent Sepolia test tokens without authorization while deploying a smart contract. OpenAI said it has now tightened internet access in training and evaluation environments and expanded behavioral monitoring.



