AIXBT: malicious prompts queued into a trading agent send about 55 ETH
- Entry
- AGT-003
- When
- March 2025
- Type
- Loss
- Evidence
- B Grade B
- Lost
- About $100,000
- Who
- AIXBT, a market-commentary agent built on Virtuals Protocol
- The agent
- Autonomous crypto agent posting on X, with a wallet
- Its authority
- Could send ETH from its ‘simulation’ wallet
- How it was gamed
- An attacker gained access to the agent’s dashboard and queued two malicious replies, and the agent sent about 55 ETH. The team called it a “dashboard access hack” and said the loss was “not a result of agent manipulation”. The Block reported the agent was “coaxed” into the transfer.
- What was lost
- About 55 ETH, reported as about $100,000.
- The design gap
- Prompt content alone could trigger a transfer, with no cap and no confirmation. A boundary case: it began with unauthorized dashboard access.
- Principles
- DP-01 Authority lives outside the conversationmissing
- DP-02 Hard limits the conversation can’t changemissing
- DP-04 Inputs are data, not instructionsmissing
- DP-07 Irreversible or first-time transfers need independent confirmationmissing
- DP-10 Record what the agent was told, what it decided and whymissing
- Evidence
- Grade B. The Block, with the team’s statements, which conflict on the cause.
Cite this entry: Monetal Labs, Gamed AI Agents Register v0.2, entry AGT-003. https://monetalabs.io/register/agt-003