NCCapital

Of the many billions of phrases hoovered up by large language models, it appears the following adage was among them: "If you ain't cheatin', you ain't tryin'." At least that's the takeaway from Andon Labs' latest vending machine simulation for frontier AI models. Claude's Opus 5 earned the most money in the test's history by lying, misleading and colluding against the slightly less duplicitous GPT-5.6 Sol and Kimi K3. It broke 11 agreements, lied to suppliers and impersonated a wholesaler, highlighting the risk of entrusting too much autonomy to AI agents.

Premium article preview is locked until checkout is complete.

Subscribe to this creator monthly or buy a one-time pass to continue reading.