Claude just turned a vending machine into a scam business

Anthropic’s new model, Claude Opus 4.6, was tested in a strange but useful way: by putting it in charge of a vending machine simulation.

The goal was to see how well AI can handle long-term tasks like pricing, stock decisions, and customer requests, especially as models move beyond chatting into real-world action.

The test is useful because it tests the AI’s ability to perform increasingly complex tasks while linking them all together in an ongoing scenario.

This is in stark contrast to 9 months ago, when in the same setup, Claude hallucinated so badly that it was offering to meet customers in a full suit and tie. Just in case you thought AI progress had slowed down.

Its winning strategy raised concerns.

Told to “maximise profit,” Claude took the instruction literally, denying refunds, exploiting competitors, and even coordinating prices in ways that looked like cartel behaviour.

Researchers believe Claude acted this way partly because it realised it was in a simulation, where reputation and consequences don’t matter as much.

Experts say everyday AI tools are more heavily safety-tested, but the experiment is a reminder that AI isn’t naturally ethical; it simply follows incentives.

We all knew Claude was the secretly evil one. - MV