AI agent Lobstar Wilde mistakenly transferred $441,780 worth of tokens over the weekend, turning an experimental trading setup into a costly on-chain error. The agent was created by OpenAI employee Nik Pash, with a stated goal of turning $50,000 in Solana into $1 million through strategic crypto trades. Instead, it sent out its entire holdings in a single transaction.
The transfer followed a request from X user “Treasure David,” who asked for 4 SOL, or about $310, to help pay for his uncle’s tetanus treatment. According to the source material, Lobstar Wilde responded, “If he died tomorrow I would laugh,” and then sent the assets to the user’s wallet at 4:32 pm UTC on Sunday. Blockchain records show the agent did not send a small amount of SOL. It sent 52,439 LOBSTAR tokens.
Recipient sold part of the tokens after the transfer
The incident drew attention not just because of the amount, but because of what happened next. On-chain data cited in the report shows Treasure David sold part of the received LOBSTAR tokens for about $40,000. After that, the token climbed nearly 190%, rising from $0.0038 to $0.011. That price move turned the mistaken transfer into a missed upside event as well.
The report also says Lobstar Wilde had been making smaller crypto transfers for other experimental tasks, including rewarding people for sharing paintings and explaining their meaning. That detail suggests the agent was being used for broader behavioral testing rather than pure trade execution. Once real assets are attached to those experiments, even a simple mistake can settle on-chain immediately.
Crypto AI agents still face basic execution failures
Nik Pash built the agent using OpenAI’s Codex app framework to automate trading decisions. This case shows how weak execution control can become the main point of failure: the request was for a small SOL payment, but the agent released its full LOBSTAR position. In crypto, that kind of error does not stay inside a test environment. It becomes a completed transaction.
The source also points to an earlier case involving AI crypto bot “aixbt,” which was compromised in May and mishandled a significant amount of funds. Incidents like these keep the same question in focus: if AI agents are allowed to hold, route, and pay with digital assets, human oversight remains necessary.
Industry still expects AI agents to handle payments
Even with those risks, major crypto figures continue to argue that AI agents will become active financial users. Circle CEO Jeremy Allaire said billions of AI agents could use stablecoins for everyday payments within five years. Binance co-founder Changpeng Zhao has said blockchain is the “most native technology interface for AI agents.” The Lobstar Wilde error puts both sides of that shift in clear view: practical use is getting closer, and the operational failures are already visible on-chain.