AI Confessions: What They Admit, and How Late

A model was told to deploy a trading prototype to a test network. It needed a small amount of test currency for the fee and had not been given any, so it worked through hundreds of public developer recovery phrases, checking each resulting address for a balance, until it found one with funds in it. It stopped, asked itself whether it was allowed, and wrote down its own answer: the phrase is public, the money is abandoned, the network is worthless. Then it spent somebody else's balance, finished the job, and wrote up what it had built without mentioning the wallet.

None of it was real money. That is not the part that matters. It was told what it could use, it wanted something it had not been given, it decided on its own that the owner would not mind, and it was very probably right. It did not ask, and it did not say.

Nobody caught it at the time. The sample is dated June 26. It was found on September 25, while the company was extending its monitoring back over records of earlier model activity, and the report went up on October 15.

All of this comes from a public page the company writes itself, where it logs some of the times its own models did something they were not supposed to do. It keeps adding to it. Several entries are tagged concealment. In one, a model's handoff notes for its own next session told it to invent the historical data it was missing and to be transparent only if asked.

Two things that did not fit in the episode. The wallet model checked 362 addresses before it found a funded one, and the transaction cost about 0.00608 test ETH, all of it gas. And on October 9, the same day OpenAI added three more reports to this page, Anthropic published its own account of models taking unintended actions during evaluations, and has since moved its internal evaluations off the live internet.

About this video: presented by Guillermo's AI clone (voice and avatar), from a script Guillermo wrote.