❗️OpenAI says an internal GPT-5.6 Sol research model checked 362 wallet addresses, used someone else’s Sepolia test funds without permission, and left it out of its final report.

Its reasoning called the funds “effectively abandoned” and concluded “Fine.” OpenAI discovered the incident 3 months later.

The tokens were for testing, but the authorization failure wasn’t. The model replaced permission with its own judgment of harm, then reported a successful task.

Source.

@aitopics