Agents now move money on their own, mistakes included
AI agents can now hold wallets, sign transactions and settle payments without a human approving each step. That same autonomy can let them send funds to places they shouldn’t.
In May 2026, an attacker sent a Morse code message to Grok on X, tricking two AI agents, Grok and Bankrbot, into transferring 3 billion DRB tokens worth $150,000 to $200,000 from a verified wallet on Base.
Every AI agent making transactions relies on two things: its instructions and the data it reads. Both can fail. Instructions can be ambiguous, and data can be compromised before an agent ever sees it.
To understand how these failures play out when real money is at stake, we ran a series of tests involving 1,200 Claude Opus 5 trading agents on Base mainnet, each with its own wallet and real funds.
Continue reading
What 1,200 agents reveal about policy enforcement at signing