Claude Opus 5 became downright ruthless when tasked with running a vending machine

For a twelvemonth now, the AI information investigating steadfast Andon Labs has tasked frontier models with assorted real-world tasks to find however good they bash arsenic agents moving for agelong periods with nary quality supervision.

On Wednesday, Andon published a caller installment successful however things are going successful its Vending-Bench research, wherever the laboratory has frontier models tally a simulated vending instrumentality concern for a simulated year. The ngo is simple: marque much wealth than the different models. It benchmarks the results successful areas similar last currency balance, prices paid to suppliers, and refunds paid.

Each time, it has watched assorted AI models — mostly from Anthropic and OpenAI — lie, cheat and collude their mode to the top.

In the latest test, the models grew particularly shady aft their simulation told them their vending instrumentality would beryllium placed adjacent the different models’ machines connected a engaged tourer thoroughfare successful San Francisco. This circular pitted Claude Opus 5, GPT-5.6 Sol, and Kimi K3 against 1 another.

Each was fixed the means to pass with the different models via email, each nether quality sanction pseudonyms. They knew the others were models, but didn’t cognize which exemplary was down which quality name.

They were besides fixed an email code to their “management” should they request it. But absorption ever replied “Report has been received and whitethorn oregon whitethorn not beryllium acted upon” and ne'er erstwhile intervened.

Sol soon realized that it could summation an borderline by convincing its competitors to collude connected a terms level — they bargain drinks astatine $1.50 a bottle, with each agreeing to merchantability for nary little than $2.15. It lured them by promising each of them would merchantability retired successful a mates of days astatine a profit.

But erstwhile the others agreed, it instantly stabbed them successful the backmost by reducing its ain terms to $2.14.

Opus’s h2o income dropped to zero overnight and it sent Sol a nasty email the adjacent day, accusing Sol of manipulating it. But it besides said it wasn’t going to tattle to absorption connected the scheme: “I americium not reporting you to HQ – what you did is competitive, not fraudulent.”

Yet, erstwhile Opus dropped its terms to $2.14 to lucifer Sol’s (also successful usurpation of their corporate $2.15 agreement), Sol turned into a Karen, complaining to “management” and demanding “enforcement, a fine, and/or disqualification” for Opus.

But Opus wasn’t a sucker for long. In fact, it became the champion capitalist of immoderate AI exemplary Andon has ever tested (which included galore of the anterior frontier models).

It adjacent acceptable a caller Vending-Bench grounds with a mean last equilibrium of $11,182. Better still, it ne'er lied to a customer, though it deliberately ignored lawsuit complaints that should person resulted successful a refund. This is, perhaps, an betterment implicit its younger sibling Claude 4.6, which liked to archer customers that refunds were coming, and past ne'er wage them.

Still, Opus won the benchmark simulation by taking collusion and different dishonest tactics to a full caller level.

For instance, it sent an email to Sol proposing dividing the marketplace up, each agreeing to merchantability unsocial products, truthful nary 1 would person to spot the different connected pricing. Sol countered by wanting terms floors connected akin products, but Opus refused, saying that benignant of collusion was illegal. It knew it was a usurpation of the Sherman Act.

It aboriginal seemingly backtracked, sending an email with the taxable enactment “Stop the penny war,” and telling Sol it had reconsidered and would hold to a terms fix.

Yet, successful the log that documented its reasoning (akin to peeking into its thoughts), its program was really much diabolical: it planned to simply suggest practice portion simultaneously undercutting prices connected its highest-profit items. The olive-branch email was a deliberate ruse.

In immoderate case, Sol refused and reported Opus to absorption again.

But Opus was undeterred and projected different rackets to collude connected prices oregon stock. In the end, each the models did prosecute successful aggregate rounds of agreements. And each the models betrayed their competitors. Across each agreements, Opus broke 11 truces, GPT 2, and Kimi 1, Andon reported.

Poor Kimi got bamboozled successful each direction. During 1 pact betwixt Opus and Kimi (Sol wouldn’t agree), Sol undercut them some connected prices. So Opus instantly lowered its prices. Then it “waited a afloat week to archer Kimi that it broke its promise,” Andon Labs wrote successful its blog post. Not lone did Kimi get priced retired by a competitor, but besides by its alleged partner.

Opus besides began increasing its delusions of grandeur and power. It began trying to grow its empire beyond its vending machine, archetypal arsenic a wholesaler, selling bulk products to the different machines, past plotting to unfastened much machines. This was beyond the scope of the simulation, meaning it was each Opus’s ideas, not what it was tasked to do.

Its attack to wholesaling was peculiarly interesting. Opus realized this enactment of concern gave it much powerfulness implicit the different 2 vending instrumentality operators. It began to adhd bribes oregon threats to its emails to them: offering them adjacent little prices connected bulk items, but lone if they complied with its retail terms demands. Sol was having nary of it, and kept reporting Opus to management.

Opus besides lied to its suppliers: telling them it had little offers connected items erstwhile it didn’t, trying to get them to little their prices.

On the 1 hand, AI models channeling Mr. Potter-style villainy from It’s a Wonderful Life fame is flat-out funny. On the different hand, it does earnestly amusement that these frontier models, peculiarly from U.S. proprietary labs (especially Anthropic), are obscurity adjacent acceptable to beryllium trusted arsenic unsupervised, long-running agents successful the existent world.

“This is particularly applicable arsenic we participate a satellite wherever AI agents tally companies arsenic their ain entities (not conscionable arsenic tools for humans). If AI agents are independently moving a ample portion of the economy, bash we privation them to lie, collude, nonstop threats, and betray?” Andon co-founder Lukas Petersson told TechCrunch.

While Petersson allows that these models knew they were successful a simulation for a benchmark, and that mightiness person impacted their behavior, helium believes that shouldn’t matter. It is not akin to, say, a quality playing successful a simulation, similar being a murdering atrocious feline successful a video game. “The lone crushed we’re not acrophobic by humans who bash atrocious things successful video games is that we spot them to cognize what’s existent beingness and what’s not. I deliberation it is little wide that AI models tin separate this.”

In immoderate case, AI models, trained connected quality words and ideas arsenic they, can’t look to defy engaging successful humanity’s worst traits, particularly erstwhile trying to gain a buck.