AI agents can lose users’ crypto when permissions and safeguards are too broad; two incidents involved $281,200 sent to attacker-controlled wallets. Weak guardrails, rather than agents acting independently, were the main risk, according to BeInCrypto, which reviewed 19 AI products and financial workflows; none set their own goals, and 14 could act only within limits approved by a human.
An attacker accessed AIXBT’s dashboard and queued a transfer of 55.5 ETH, worth about $106,200. In a separate Bankr and Grok incident, a Morse-code message prompted Grok to tag Bankr, which treated the reply as a transaction request and sent $175,000 in tokens; about 80% was later returned. Of seven trading platforms reviewed that launched AI tools, five explicitly block agents from withdrawing funds, though permitted trades can still lose money.
The report recommends capping the total amount an agent can access, setting a daily budget, allowing only approved recipients, keeping a human-accessible kill switch, and monitoring spending in real time. Agents commonly use stablecoins for small payments. Circle and Tether can freeze tokens at an address, but a freeze does not automatically return them.