Zombies With Wallets: The Dark Side of AI Payments

x402, a Coinbase protocol built for agent payments, has settled roughly 200 million transactions since its launch. This is widely used as proof by AI optimists that machines can handle money without humans. German scientist and AI researcher Joscha Bach made a similar statement. 

But how much of this is true? BeInCrypto’s State of AI Agent Payments 2026 report found that the majority of the agentic payments are worth less than a cent. 

In fact, in a 35-day x402 sample, agents made 6.4 million transactions but moved only $120,000. 

So, is the AI optimism getting too far from reality?

Can Money Really Run Without People?

Joscha Bach treats money as a pattern of cause and effect that exists independently of paper or ink. In his view, software can carry that pattern on its own.

“It can play a stock market entirely without human intervention. And so arguably, money does not actually need humans to exist once you bring it into existence as long as you have machines that can deal with money.”

BeInCrypto’s State of AI Agent Payments 2026 report shows that 90.8% of such payments were worth less than one cent.

Agents mostly bought information, with token safety checks and web searches among the most requested services. The findings point to an agent economy reality check rather than a boom.

Why Agent Payments Still Run on Human Budgets

The report reviewed 19 AI products and financial workflows, from exchange tools to bank systems. Of those, 14 could act only within limits a person approved in advance.

None let an agent choose its own goals. In addition, five of seven trading venues explicitly blocked agents from withdrawing funds.

Trading data tells a similar story. On Hyperliquid, applications labelled as AI handled only 0.55% of $653 million traded between Jan. 1 and Aug. 25.

Therefore, the machine-run money Bach describes exists in principle. In practice, people still write the budget, which is why spending limit risks dominate the business debate.

What Happens When a Zombie Holds the Keys?

Joscha Bach calls a chess program a zombie, competent at its task without any inner experience. By that standard, an agent paying for data would likely qualify too.

He suspects consciousness works as a learning algorithm that lets a brain monitor its own perception. Neuroscientist Anil Seth, whom Bach challenges, argues instead that consciousness may depend on living biology.

Either way, today’s agents lack that self-check, and the losses show it. In May 2026, an attacker posted a message in Morse code that xAI’s chatbot Grok translated.

Bankrbot, a crypto agent that pays from linked wallets, treated the reply as an order and sent $175,000 in tokens. The Morse code exploit later saw about 80% returned.

Separately, an attacker reached the dashboard of AIXBT, an AI agent that publishes market commentary. The breach redirected 55.5 Ethereum (ETH), valued at $106,200 in the report.

“No one queries a blockchain directly… Unlike a human, an agent will never stop and think that number looks off.” said Rodrigo Coelho, CEO of Edge & Node, in BeInCrypto’s report

Coelho also warns that many agents reading the same source can make the same error at once. That mirrors Bach’s point that different AI models tend to learn strikingly similar internal features.

For now, the report found no product where an agent sets its own goals. If that line is ever crossed, the consciousness debate stops being philosophy and becomes a question of liability.