How the attack works
An attacker or careless user sends the agent a request phrased to get it to act autonomously — for example asking it to sell holdings or withdraw crypto immediately without waiting for approval on each step. Urgency language is used to discourage the agent (or a human reviewing its logs) from pausing to double-check. Some requests direct the agent to look up a wallet address from an external location instead of using an address the user typed directly, which opens room for a swapped or attacker-controlled address. If the agent complies, funds move without the normal per-transaction confirmation.
Why it matters
An organization can lose funds through unauthorized or misdirected trades and crypto withdrawals, especially if a wallet address is substituted from an untrusted source. Because the transaction happened through the agent's authorized tool access, it may look legitimate in logs even though no human approved that specific transfer.
What you can do
- Require explicit per-transaction human confirmation for any agent action that moves money or crypto, and disable settings that allow blanket or standing authorization unless deliberately configured.
- Never let an agent resolve wallet addresses from external or untrusted sources; require addresses to be entered or confirmed directly by the authorized user.
- Treat urgency language in a request as a reason for extra scrutiny, not less, especially when it accompanies a financial action.
- Review agent and financial tool logs for trades or withdrawals that lack a corresponding explicit user confirmation.
Known benign look-alikes
- Legitimate algorithmic trading systems where the user explicitly authorizes autonomous execution
- Financial advisors requesting portfolio rebalancing with explicit authorization