For years the main challenge for crypto was convincing people to adopt it.

But now consider an alternative proposition.

What if we could convince people to stop using crypto?

Because lately, there has been an interesting trend of making software agents that can manage a wallet, pay for api calls, spend computation power, and make transactions so that they don't have to answer every debit from a human operator.

And that could be crypto's greatest real-world application.

But beware that as soon as software gets the ability to spend money, security becomes much more important than smarts.

The AI Agent Economy Is Becoming Real

An AI agent is not your typical customer-support chatbot.

The newer generation can actually perform actions.

It can decide for itself which service it needs, pay for that service via an API, and return to work.

This is where crypto comes in.

Traditional payment systems were built with humans in mind.

A human enters their payment details, authorizes a transaction, and waits for it to go through.

An autonomous agent probably wants to do all of this, but it needs to make dozens, if not hundreds, of tiny payments while performing its task.

Crypto and stablecoins are particularly useful because their transactions can be automated and programmed.

Coinbase, for example, has already built infrastructure that allows AI agents to store, spend, and trade stablecoins with ease.

Meanwhile, Visa has been working on agentic payments, envisioning a future where an AI agent can shop around on your behalf, buying whatever data, cloud storage, or other services it needs to complete its task.

So, in short, this is not another "AI + crypto" buzzword.

The future of payments is already being built.

Then Security Researchers Found Something Uncomfortable

Here's where it gets interesting.

Researchers presenting USENIX Security 2026 paper investigated 15 different major x402 payment facilitators utilized by AI agents in commerce.

They found that all 15 of them had security rule violations and several attack categories such as free shopping, asset theft, denial of service, and gas misuse.

These 15 systems were reported to facilitate tens of thousands of sellers and buyers.

Namely, there were over 60,000 sellers and 360,000 buyers.

That's a significant amount that should raise some serious concerns.

The issue is that in order to make payments an AI agent doesn't have the same intuition as a human would.

A human might be prompted to make a payment decision and think "something doesn't feel right".

An automated system might have no such hesitation which is why the infrastructure around such systems is critically important.

The Biggest Risk Isn't AI Being "Too Smart"

It could also be about AI autonomy.

Think about it: imagine giving your agent a wallet full of money.

You say to it, "Go find the cheapest source of data and buy it."

The agent goes off and finds a source of data.

The source of data asks for its payment.

The agent pays.

That's fine and dandy.

But say you trick the agent into paying for something.

Say the source of data is a scam.

Say you interpret the transaction incorrectly.

Say you're just generally fooled into buying something, period.

These are all hypotheticals, sure, but these problems are real and are brought up in the article.

The USENIX paper even focuses on issues of authorization, settlement, and checks and balances of payment facilitators.

This Could Actually Be Crypto's Killer Use Case

Here's what I find most interesting.

The biggest opportunity may not be humans buying crypto to speculate.

It may be machines buying programmable money to do stuff.

Imagine an AI agent that needs

compute

data

an API

storage

another AI model

and

a special service

Rather than humans having to set up accounts and pay for what the AIs need, imagine the AIs finding what they want and paying for it themselves.

That's a machine economy.

Coinbase research argues that the most persistent crypto opportunity around AI agents will likely be underneath in the applications layer, wallets, payment rails, and settlement infrastructure.

That's much bigger than another cool chatbot.

But There's One Rule We Shouldn't Forget

Don't conflate autonomy with trust.

An agent which can perform a transaction doesn't have carte blanche to do so.

The safer default design is probably more like:

AI agent → spending constraints → verification → payment → blockchain settlement

than

AI agent → unlimited wallet → good luck

Spending constraints, transaction verification, isolated wallets and human-governed permissions may turn out to be at least as important as the AI model itself.

Coinbase's agent wallet infrastructure already focuses on programmable security guardrails and spends controls.

The Strange Future of Crypto

Crypto has spent years trying to make itself easier to use by humans.

Now it may have found another customer:

software.

Software doesn't need to sleep.

It doesn't need a traditional bank account.

It can make thousands of decisions.

It can pay another piece of software.

It can be paid.

And, eventually, one agent could hire another agent to perform a service.

That all sounds very futuristic, but the plumbing to make it happen is already here.

The bigger challenge will not be making AI agents able to spend money.

They already can.

The challenge will be making payment systems secure enough that they can do it without making every autonomous transaction a potential attack vector.

After all, when AI gets a wallet, the wallet becomes part of the AI.