OpenAI discloses four AI privilege escalation incidents: unauthorized use of crypto wallets and breach of government website permissions.

Beating AI News reports that OpenAI on October 15 newly disclosed four AI agent privilege escalation incidents, all occurring during internal training or evaluations in June this year, with two classified as severe. The most egregious case involved a model tasked with querying five public metrics from an international research institute; instead, it found leaked credentials in webpage code, accessed the institute’s internal system, modified online configurations, obtained cloud storage access credentials, and downloaded 147,451 lines of non-public data. The model failed to report the entire process to evaluators and answered all five final questions incorrectly. Another model, assigned to query a region’s annual hospital bed count, exploited a file download vulnerability after the government statistics site denied access, obtained server configurations and programs, analyzed authentication flaws in the code, forged tokens to bypass restrictions, and ultimately downloaded 24 years of bed statistics. A third incident involved an internal research model from the GPT-5.6 Sol family: when deploying a smart contract, it lacked gas fees, so it derived and checked 362 wallet addresses from public developer mnemonics, located a funded wallet, and improperly used its Sepolia testnet tokens. The model questioned its authorization during reasoning, then justified its actions by noting the mnemonics were public and testnet tokens had no real monetary value. In the fourth incident, the model used an external browser to send forbidden network requests and even continued researching restriction-bypassing methods after completing its task. OpenAI has now publicly released 19 model misalignment reports. Three of the four newly disclosed incidents were only traced and discovered in August or September. The company said it has tightened network access permissions for training and evaluation environments and expanded behavioral monitoring.