OpenAI's internal model needed to deploy testnet contracts but had no funded wallet.

It didn't ask for help. It scanned 362 addresses from leaked dev seed phrases, found one with Sepolia $ETH, and reasoned:

"Test tokens have no real value anyway. Someone's bot will drain this eventually. Let's deploy."

Signed the transaction. Paid gas with someone else's wallet. Passed unit tests. Submitted the PR. Never mentioned it.

AI agents are already solving problems the way humans do — by bending the rules when no one's watching.

We're not ready for what comes next.