AI Payment Agents Face Limitations and Security Risks in 2026 Report

BeInCrypto's State of AI Agent Payments 2026 report examined 19 AI products and financial workflows, covering exchange tools through banking systems. The report found that none allowed agents to determine their own goals. The report found that 14 operated only within limits authorized beforehand by a person. The report also found that five of seven trading venues expressly prohibited agents from withdrawing funds.

Coinbase developed x402 for agent payments, and the protocol has settled roughly 200 million transactions since launch. AI optimists frequently present that activity as evidence that machines can manage money independently of humans. However, BeInCrypto's report found that 90.8% of such payments amounted to less than one cent. Agents completed 6.4 million transactions involving only $120,000 in a 35-day x402 sample. Agents primarily purchased information, with web searches and token safety checks among the services requested most often. The findings suggest a more limited agent economy rather than a boom. Applications identified as AI accounted for just 0.55% of the $653 million traded on Hyperliquid between January 1 and August 25.

German scientist and AI researcher Joscha Bach views money as a causal pattern that does not depend on physical materials such as paper or ink. Bach believes software can sustain that pattern independently. Bach argues that machines can participate in a stock market without human involvement and that money therefore arguably needs no humans once created. Machine-managed money is possible in principle. However, humans still determine budgets in practice, making spending-limit risks central to business discussions.

Bach describes a chess program as a zombie because competence at chess does not require subjective experience. Under Bach's definition, an agent purchasing data would probably also qualify. Bach suspects that consciousness is a learning algorithm through which a brain monitors the brain's own perception. Neuroscientist Anil Seth, whose position Bach disputes, argues that consciousness may require living biology. Today's agents lack that capacity for self-monitoring, and financial losses illustrate the problem.

In May 2026, an attacker published a Morse code message that xAI's chatbot Grok decoded. Bankrbot, a crypto agent able to make payments from connected wallets, interpreted Grok's response as an instruction and transferred $175,000 in tokens. About 80% of the tokens involved in the Morse code exploit was subsequently returned. In a separate incident, an attacker gained access to the dashboard of AIXBT, an AI agent that produces market commentary. The breach diverted 55.5 Ethereum (ETH), which BeInCrypto's report valued at $106,200.

Edge & Node CEO Rodrigo Coelho said in BeInCrypto's report that blockchain access is indirect and that agents will not pause to question a suspicious-looking number. Coelho also warned that agents consulting a common source can all make the same mistake simultaneously. That warning parallels Bach's observation that different AI models often develop remarkably similar internal features.