Buy
Market
🔥
Prediction Market

5 Safety Rules for Letting an AI Agent Access Your Crypto

AI agents can make payments but pose risks. Experts advise capping amounts, setting budgets, and using allowlists to avoid losses.

18/11/2026 05:2617 min read

AI agents can now make payments on a user’s behalf for tasks like booking services, purchasing data, or executing trades without manual approval. That capability has sparked concerns among crypto users about the risk of an AI agent draining their wallet.

This topic was explored during BeInCrypto’s latest Market Intelligence Experts Council episode, where Senior Research Analyst Michael Thoma and Dune Head of Research Filippo Armani examined the report The State of AI Agent Payments 2026.

The quick answer is that an AI agent can cause financial loss, but that rarely stems from it “going rogue.”

Understanding AI Agent Payments

Thoma compared an AI agent to a digital personal assistant holding a prepaid debit card.

“You give it a task, and you give it permissions and a rule set, so it can go out, complete the task for you and make a payment on your behalf.”

These agents typically pay with stablecoins since card networks are not designed for micropayments. Blockchain technology allows agents to settle small sums almost instantly.

However, not all automated payments are driven by AI agents. Armani noted that onchain, agents and regular trading bots can appear similar, with numerous small payments occurring continuously. Armani stated:

“When you see many small payments on average, and constant throughout the time, it’s probably bot activity, but it’s not necessarily an AI agent.”

He pointed out that the difference lies in the infrastructure. Agent payments operate via dedicated protocols like Coinbase’s x402 and Tempo’s MPP, while bots do not use these rails.

Main Danger: Inadequate Safeguards, Not Rogue Software

Thoma said that incidents so far have not resulted from agents acting autonomously.

“This is software, and software can have bugs. A lot of it is not the agent going rogue… It’s just that the guardrails put in place were not specific enough or robust enough.”

The BeInCrypto report documents two cases where $281,200 was moved to attacker-controlled wallets before any funds were retrieved:

  • AIXBT: An attacker accessed the agent’s dashboard and initiated a transfer of 55.5 ETH, valued at roughly $106,200. Control over the dashboard translated into control over the funds.
  • Bankr and Grok: An attacker submitted a message encoded in Morse code. Grok translated it into a reply that mentioned Bankr, and Bankr interpreted that reply as a transaction request, sending tokens worth $175,000. Approximately 80% was later refunded.

In both scenarios, the agent operated exactly within its permitted settings. The issue was what those settings allowed.

How Much Autonomy Do Agents Actually Have?

Less than media reports suggest. BeInCrypto examined 19 AI products and financial workflows. None set their own objectives, and 14 could act only within bounds pre-approved by a human.

Exchanges also exercise caution. Among seven trading platforms reviewed that introduced AI trading features, five explicitly prevent agents from withdrawing funds.

That safeguard has a limitation, though. An agent restricted from withdrawals can still lose money via the trades it is authorized to execute.

Five Guidelines Before Granting an Agent Access to Funds

The BeInCrypto report outlines five controls addressing various risks. Thoma emphasized the first three during the conversation.

  1. Limit the total amount. “You don’t want to give an agent all your money,” Thoma said. “Just give it a set amount, a small amount that you’re willing to entrust to this very new technology.”
  2. Establish a daily budget. A single large payment is easy to detect. Many small payments can accumulate unnoticed.
  3. Permit only pre-approved recipients. An allowlist means the agent “can never run off and pay your nemesis… It can only go to pre-approved addresses,” Thoma explained. Rodrigo Coelho, CEO of Edge & Node, called this the most valuable control in the report, “because most real losses are not the wrong amount, they are the right amount going to the wrong place.”
  4. Maintain an accessible kill switch. A method to immediately halt the agent’s spending is necessary. “A kill switch has to be immediate and reachable by a human, otherwise it is decorative,” Coelho said.
  5. Monitor spending in real time. A live record of payments highlights problems while they remain manageable.

No Guarantee of Recovery

Stablecoin issuers like Circle and Tether can freeze tokens at a specific address, and they do so frequently. However, a freeze only stops further movement of funds; it does not automatically return them to the owner.

AI Agent Crypto Wallet: Key Takeaway

The technology functions but remains in early stages. The safest strategy is to restrict how much an agent can access from the start. As Will Papper, Director of Product for Agent Payments at Cloudflare, noted in the report:

“If an agent is responsible for $10, you can worry less about its spending than if it is responsible for $1,000.”

Or, echoing a classic investing principle: never entrust an agent with money you cannot afford to lose.

Share to

Disclaimer: this article comes from third-party media and is provided for reference only. It does not constitute investment advice. Crypto and other financial products carry significant price volatility risk, so please make your own decisions carefully.

Related articles