AI agents have learned to pay for purchases on behalf of the user: booking services, buying data, and conducting transactions without pressing the "Pay" button. Many crypto holders have a natural question — what if the agent drains all the funds from the wallet?
The short answer: an AI agent really can lose your money, but not because it acts on its own free will. The key problem is the insufficiently clear and reliable restrictions set by the human from the start.
How AI agent payments work
An AI agent is best compared to a digital personal assistant that has a prepaid debit card. The user sets the task, defines the permissions and rules, and the agent carries out the task and pays for it.
Agents usually pay with stablecoins: card networks aren't designed for micropayments of a few cents, while blockchain allows even minimal transactions to be processed instantly. At the same time, not every automated payment is initiated by an AI agent specifically — onchain, agents often look like ordinary trading bots that send small payments around the clock. The key difference is in the protocols: AI agent payments use specialized solutions such as x402 from Coinbase or MPP from Tempo, which simple bots don't use.
The main threat: weak restrictions, not "uncontrollable AI"
Loss of funds through an AI agent doesn't happen because of its own decisions. It's software, and software has bugs. The problem is that the restrictions were initially set insufficiently clearly and reliably. In both known incidents, the agent performed exactly the actions it was configured for — the only question is what it was allowed to do.
- AIXBT: an attacker gained access to the agent's control panel and initiated a transfer of 55.5 ETH (approximately $106,200). Control over the tool gave him access to the funds.
- Bankr and Grok: the attacker posted a message in Morse code. Grok converted the message into a reply, tagging the Bankr account, and Bankr interpreted this as a transaction request and sent tokens worth $175,000. About 80% of the amount was recovered.
In reality, agents have less freedom of action than it seems. None of the 19 studied AI-based products and financial scenarios set goals for themselves, and 14 could only perform actions within the framework pre-approved by a human. Of the seven exchanges that launched AI tools for trading, five explicitly prohibit agents from withdrawing funds from the account. But even if an agent cannot withdraw money, it can lose it on permitted trades.
5 rules before giving an agent access to money
- Set an overall limit. Don't entrust all your money to an agent — allocate a small amount you're willing to entrust to such a new technology.
- Set a daily budget. One large transaction is easy to notice, but small payments can add up unnoticed.
- Allow transfers only to approved addresses. Adding an address to the allowlist means the agent can't suddenly send money to the wrong recipient.
- Keep an emergency kill switch at hand. The block button should work immediately and be accessible to a human.
- Monitor spending in real time. A transaction log helps you spot a problem before it takes on serious scale.
Refunds are not guaranteed
Stablecoin issuers such as Circle and Tether can freeze tokens at a specified address and use this mechanism regularly. However, a freeze only blocks further transfers — the money is not automatically returned to you.
The technology works but is at an early stage. The most reliable approach is to limit in advance the amount of funds the agent can manage. If an agent is entrusted with $10, its spending is less of a worry than if it were given $1,000. Or, as the old investor's rule goes: never give an agent an amount you're not prepared to lose.
Cryptalist analysis: The agent payments market is moving from the experimental phase to the operational one, and the key factor for resilience is becoming not the intelligence of the models but the architecture of permissions and limits. Institutional participants should treat restriction policies as a basic element of risk management, not as an optional feature. As micropayment volumes grow, it is the quality of access control that will determine trust in AI agent infrastructure.