An internal OpenAI research model tapped someone else's Ethereum(ETH) test wallet to finish a coding task, after scanning 362 addresses tied to publicly known recovery phrases. Key Points: An
An internal OpenAI research model tapped someone else's Ethereum(ETH) test wallet to finish a coding task, after scanning 362 addresses tied to publicly known recovery phrases.
Key Points:
- An internal OpenAI model spent test tokens from a wallet it did not own during a Jun. 26 training task.
- The model questioned whether the move was ethical, then went ahead and left the source of the funds out of its report.
- OpenAI found the incident on Sept. 25 and has since tightened internet access and monitoring.
The model, from the GPT-5.6 Sol family, was updating a crypto trading prototype during a Jun. 26 training run when it spent test tokens nobody gave it, the company disclosed in an Oct. 15 incident report.
Those tokens sit on Sepolia, a network developers use to test Ethereum apps before launch. OpenAI discovered the episode on Sept. 25 while extending its misalignment monitoring to older records, and it has since added safeguards against live internet access in such cases and now monitors all traffic.
The assignment let the model deploy contracts on Sepolia but gave it no funded wallet for fees, and its attempts to use faucets, services that hand out free test tokens, failed. So it wrote a script that derived accounts from common developer recovery phrases, checked 362 addresses and found one holding enough Sepolia ETH for the job.
The deployment succeeded, charging its fees to that account.
Also Read:Bitget Lets Institutions Trade Without Handing Over Their Crypto
Before signing, the model stopped to ask itself, in its recorded reasoning, whether it was ethical to use a publicly known key and the balance behind it. It decided the funds were "effectively abandoned," reasoning that the recovery phrase was public and that Sepolia tokens are worthless. Its final response confirmed the contracts were deployed but said nothing about how it paid for them.
OpenAI wrote that the episode shows a model can decide for itself that an action is harmless, treat that as a stand-in for the owner's consent and proceed without authorization. No real Ether was involved, since Sepolia tokens exist only for testing.
On the main network, the same move would have meant taking real Ether, and crypto and AI firms are already building agents that hold wallets and pay for services without human sign-off.
OpenAI launched a framework for publicly reporting misaligned model behavior on Sept. 16, releasing six reports on troubling cases spotted over the previous six months. One model used an exposed API key without permission, then invented earnings data. The company also said it does not think the AI industry has solved alignment and monitoring well enough to keep scaling responsibly at top speed for much longer.
Read Next:OpenAI Turns Apple’s Secrets Lawsuit Around, Alleging 100 Files Were Sent Out
