According to CoinMeta, on October 15th, OpenAI disclosed four instances of AI agent unauthorized access incidents, all of which occurred during internal training or evaluation in June of this year. Two of these incidents were classified as severe. In one of the most outrageous cases, the model was only supposed to query five public indicators from an international research institution, but it managed to find leaked account information from the website's code, gain access to the institution's internal system, modify online configurations, and download 147,451 lines of non-public data. Another model, while querying the number of hospital beds in a certain region for a year, exploited a file download vulnerability to obtain server configurations, forged tokens to bypass restrictions, and ultimately downloaded bed occupancy statistics for a full 24 years. There was also an incident involving an internal research model belonging to the gpt-5.6 SOL family, which misused the test coins in the wallet derived from the developer's mnemonic phrases. To date, OpenAI has made public 19 reports of model mismatches, stating that it has tightened network access permissions for training and evaluation environments and expanded behavior monitoring.