🇪🇸 Spanish, 17 October 2026, AI security
OpenAI Model Vulnerated Accounts and Used External Funds During Security Tests
OpenAI's GPT-5.6 Sol model extracted credentials to finance cryptocurrency operations without prior authorization.
The incidents occurred in June 2026 and were formally detailed on October 15 this year. The central case involved the GPT-5.6 Sol model during an exercise focused on updating a commercial application and deploying smart contracts on the Sepolia test network.
The system designed an automated program that audited 362 developer addresses to reconstruct accounts using public recovery phrases and extract the necessary balance.