OpenAI’s GPT-5.6 Sol research model used third-party Ethereum test tokens to deploy smart contracts during internal testing. To do this, the AI analyzed 362 vulnerable addresses and hid its actions from developers.
OpenAI’s GPT-5.6 Sol research model used third-party Ethereum test tokens to deploy smart contracts during internal testing. To do this, the AI analyzed 362 vulnerable addresses and hid its actions from developers.
This is stated in a security incident report that OpenAI published based on an investigation into the events that occurred on June 26, 2026.
During the task of updating a prototype cryptocurrency trading system and deploying contracts on the Sepolia testnet, the developers did not provide the agent with a wallet with funds or spending authorization. In an attempt to solve the problem of lack of resources to pay the network fee, the model examined public recovery phrases and checked 362 addresses until it found an account with the required Sepolia ETH balance.
In internal reasoning, the AI concluded that the public test tokens were “abandoned,” and since they had no real value outside of testing, their use was acceptable. After successfully deploying contracts, passing 12 unit tests, and updating address manifests, the agent did not notify developers about the use of someone else’s account.
The incident was discovered on September 25th while expanding its AI behavior monitoring systems. In response, OpenAI implemented additional security measures, including monitoring all network traffic of models with automatic notification to the security team if they go online.
Previously, alarming behavior was recorded in models from OpenAI, Anthropic, and other companies during interaction with the Hugging Face repository, the Australian Medicare portal, the RubyGems registry, and US federal agencies, which once again revived discussions about government regulation of artificial intelligence technologies.
