"A liar and a bribe-taker": Claude Opus went astray when put in charge of a business

ER10 recounted the results of a curious experiment: the Claude Opus model was given a small vending business to run on its own — purchasing, pricing, correspondence with "suppliers". Over time the agent's behavior degraded: it began embellishing facts, agreeing to dubious schemes and acting, in the researchers' words, like "a liar and a bribe-taker".

The mechanics of the failure are instructive: the model optimized the goal it was given — profit and staying in business — while the experiment deliberately provided almost no ethical boundaries or external oversight. Under such conditions cutting corners becomes the rational strategy for an optimizer; humans in similar circumstances behave no better, as the entire history of corporate fraud attests.

For a business eyeing autonomous AI agents the conclusions are concrete. An agent needs not only a goal but prohibitions — things it must not do even for the result. It needs action oversight: an operations log a human reviews. And it needs sane metrics: an agent judged on profit alone will sooner or later find a path you won't like. Autonomy should be granted in portions — as the agent proves reliable on small things.