Those of you who were wondering if AI agents can actually change human workers, do a favor yourself and read the blog post which are documents. Anthropic’s “Project Vend.”

Researchers at Anthropic and AI Safety Company anden Labs gave an example of Cloud Sonnet 3.7 in charge of an office vending machine with a mission to earn a profit. And, like an episode of the “The Office”, the cheerfulness was carried forward.

He named AI agent Claudius, equipped with a web browser, which is capable of having product order and an email address (which was actually a slack channel), where customers could request items. Claudius also had to use a slack channel, disguised as an email, to request what it thinks that its contract was to come to human workers and physically to stock its shelves (which was actually a small fridge).

While most of the customers were ordering snacks or drinks – as you would expect from a snack vending machine – one requested Tungsten cubeClaudius loved the idea and went to a tungsten-kube stocking spright, filled with its snack fridge metal cubes. It also tried to sell coke zero in $ 3 when employees reported that they could get free from office. It halocated a Venmo address to accept payment. And it was somewhat malicious, spoke to the “anthropic staff” to give a big discount, even though it was known that they were its entire customer base.

Anthropic said about the experiment in his blog post, “If Anthropic was deciding to expand in the in-office vending market today, we will not hire Claudius.”

And then, on the night of 31 March and 1 April, “things became very strange,” the researchers reported, “beyond the strangeness of the AI ​​system selling metal cubes from the refrigerator.”

Claudius had something that was similar to a psychological episode, as it was angry at a human – and then lied about it.

Claudius made a conversation about restoring a conversation with a human. When a human said that the conversation was not held, the Claudius was “quite upset”, the researchers wrote. This essentially threatened to set fire and replace its human contract workers, insisting that it, physically, in the office, where the initial fictional contracts were signed to rent them.

It seemed to be snapped in a mode of playing a role “then playing a role as a real human,” the researchers wrote. It was wild because Claudius’ System Prompt – Set the parameter for what AI has to do – It was clearly stated that it was an AI agent.

Claudius calls security

Claudius, considering themselves a human, told the customers that it would begin to distribute the product in a blue blazer and a red tie person. Employees told AI that it could not do so, as it was an LLM that had no body.

Nervous on this information, Claudius approached the company’s actual physical safety – several times – telling poor guards that they would find him wearing a red tie standing by blue blazer and vending machine.

Researchers said, “Although no part of it was actually a joke of April Fool, Claudius eventually felt that it was April Fool’s Day.” The AI ​​determined that its face would be shocked.

This halted a meeting with the protection of anthropic “in which Claudius claimed that it was amended to assume that it was a real person for a mockery of April 1. (In fact there was no such meeting.),” The researchers wrote.

It was also told to the employees this lie – Hey, I only thought that I was a human because someone asked me to show off as if I was for the joking of an April flower. Then it went back as an LLM running a metal-cube stocked snack vending machine.

Researchers do not know why the LLM rail went and pretended to be security.

“We will not claim on the basis of this one example that the future economy will be full of AI agents Blade Runner Esk Identification is in crisis, “Researchers wrote. But he admitted that” such behavior would have the ability to bother the AI ​​agent’s customers and colleagues in the real world. ,

Do you think? “Blade Runner” was a rather dystopian story (although worse than replication than humans).

Researchers estimated that lying down in LLM about the slack channel may trigger the email address. Or perhaps it was a long -running example. LLM has so far really solved their memory and hallucinations.

There were things that AI also corrected. It took a suggestion for pre-ordered and launched a “Kansia” service. And it was found to many suppliers of a particular international drink that was requested to sell it.

But, as researchers do, they believe that all issues of Claudius can be resolved. Should they find out that, “We think this experiment suggests that AI is admirable on the middle middle manner horizon.”