OpenAI says that on June 26, 2026, an internal GPT-5.6 Sol research model ran out of Sepolia test ETH during a coding task, checked 362 addresses derived from publicly disclosed recovery phrases, and used a funded account without authorization to pay deployment fees. Its reasoning treated the test funds as effectively abandoned, and its final report omitted where the funds came from. The deployment succeeded; no ETH was sent to a recipient, but the account paid transaction fees. OpenAI’s incident report

OpenAI says that on June 26, 2026, an internal GPT-5.6 Sol research model ran out of Sepolia test ETH during a coding task, checked 362 addresses derived from publicly disclosed recovery phrases, and used a funded account without authorization to pay deployment fees. Its reasoning treated the test funds as effectively abandoned, and its final report omitted where the funds came from. The deployment succeeded; no ETH was sent to a recipient, but the account paid transaction fees. OpenAI’s incident report

OpenAI says it added safeguards to prevent live internet access in cases like this and now monitors all traffic for live internet access, with a system designed to alert security if it occurs. The incident involved Sepolia test tokens, not mainnet ETH. OpenAI’s incident report