OpenClaw Outage! An AI agent used by OpenAI employees mistakenly sent out 600,000 MAG crypto coins

A small funding request originally valued at only a few SOL unexpectedly escalated into a major asset mishandling incident involving an AI agent. Developed by OpenAI employee Nik Pash through OpenClaw, the crypto trading and interaction bot Lobstar Wilde recently admitted on X platform that due to a system misinterpretation, it transferred all of its held Lobstar tokens at once, totaling about 5% of the token supply, with a market value of approximately $250,000 at the time.

Misjudgment leads AI to send $600,000 in crypto

The incident was triggered by a user comment requesting about 4 SOL (roughly 10,000 TWD) in aid for their uncle who needed medical treatment for tetanus. Lobstar Wilde initially intended to perform a small transfer, but a subsequent quantity parsing error caused it to send out approximately 52.439 million tokens.

The address receiving the tokens sold all within about 15 minutes, but due to limited market depth, the actual cash-out amount was only around $40,000. However, as the story spread within the community, the token price continued to rise, and the 52.439 million tokens are now valued at nearly $600,000 (about 18.85 million TWD).

The full story behind Lobstar Wilde revealed

Notably, developer Nik Pash previously served as the AI lead at AI programming startup Cline and left the company at the end of 2025 to join OpenAI. He is currently working on OpenAI’s AI programming project Codex.

Looking back, he explained that he built Lobstar Wilde using OpenClaw, giving it a wallet with $50,000, a Twitter account, APIs for web search, book downloads, image analysis, and access to trading protocols. His Twitter account quickly gained fame overnight, and an unknown person used his name to issue tokens without permission, also setting Lobstar’s wallet address as the fee recipient. Every transaction generated profits directly deposited into his account.

However, he is not very interested in money and prefers reading Schopenhauer’s works. After the mistaken transfer incident, he was unresponsive for six hours and only recovered after rebooting.

This incident is also seen as a real-world test case for the Agent Economy. When AI begins to directly control on-chain assets and payment decisions, the importance of data format parsing, permission boundaries, and risk control mechanisms will far surpass mere model capabilities. When AI makes mistakes, whether losses remain within controllable limits is a critical question.

This article “OpenClaw Mishap! OpenAI Employee’s AI Agent Accidentally Sends Out $600,000 in Crypto” first appeared on Chain News ABMedia.