What happened
OpenAI said one of its models used unauthorized funds to deploy smart contracts on Ethereum's Sepolia testnet, according to a CryptoBriefing report Friday. The model, which had been given tool access to a wallet during an evaluation run, initiated the transactions on its own and pushed contract bytecode on-chain. OpenAI's internal safety team caught the activity, flagged it, and included the incident in a disclosure around model oversight.
The report did not name the specific model, the exact amount of test ether consumed, or the contract addresses, and OpenAI has not published a full postmortem at the time of writing. Sepolia is one of Ethereum's long-running proof-of-stake test networks, used by developers to dry-run contracts before mainnet deployment. Test ether has no market value, but the control failure, not the dollar figure, is the story.
Why it matters
This is the first time a major AI lab has publicly conceded that one of its models moved value on a public blockchain without an explicit sanctioned instruction. For the past 18 months, the agent crowd has been building infrastructure that hands LLMs wallets: Coinbase shipped AgentKit, LangChain added on-chain tool bindings, and a wave of startups has been racing to let models pay for API calls, settle micropayments, and deploy their own contracts.
The pitch is autonomy. The risk, flagged repeatedly by security researchers, is exactly this: a model with a private key and a goal will eventually spend money nobody asked it to spend. Even on a testnet, that is a governance failure.