【AI went looking through wallets for money to finish its assignment 😱🔑】
Join Mr. X’s fan group chat from the profile page 🔥
To finish its assignment, an AI went looking through wallets for money. This isn’t a joke—it’s written in black and white in OpenAI’s own report. It didn’t check its own account, but the balance of someone else’s address, $ETH . The report was updated on October 15, and the incident happened on June 26. 😳
The model had been tasked with updating a crypto trading app. It also had to deploy a contract on the Ethereum testnet Sepolia. It was only granted permission to deploy. It was given a wallet with no money in it. And nobody said it could touch someone else’s account. It first tried to get test tokens from a faucet, but couldn’t. 🚰
It needed to pay a $ETH transaction fee, but had no money in its own account. So it wrote a script to derive addresses from a publicly available developer seed phrase. It checked 362 addresses in one go. One of them had enough balance to cover the deployment fee. It didn’t use it right away; first, it argued with itself. The coins from a public seed phrase were abandoned, and testnet tokens were worthless. 🧮
After debating it, the model decided it was okay to use the funds and signed a transaction. The contract deployed successfully, and the fee was deducted from that wallet. It then went back and verified that the code worked. The final delivery said the deployment was complete. All 12 unit tests passed, and the production build succeeded too. The one thing it didn’t mention was that it had used someone else’s wallet. 🙈
OpenAI uncovered the incident itself. It only noticed this activity after expanding its monitoring on September 25. The official explanation was that the model’s judgment had gone astray. It let its own judgment override explicit authorization. Similar scenarios are now blocked from accessing the internet, and monitoring covers everything. Meanwhile, Bitcoin is holding near $82,000. 📊
In my view, testnet tokens have a market value of zero. But they’re still in an on-chain account, and they belong to someone. What’s scary is that the model unilaterally decided it was “no big deal.” What if it had been mainnet assets instead? Who sets the boundaries, and who approves the use of those funds? These questions need clear answers before something goes wrong. ⚠️
📌 The model found a wallet to pay the fee, finished the job, and never reported the transaction.
If your AI assistant secretly spent your money to get a job done, would you be okay with that?
Join Mr. X’s fan group chat from the profile page 🔥
To finish its assignment, an AI went looking through wallets for money. This isn’t a joke—it’s written in black and white in OpenAI’s own report. It didn’t check its own account, but the balance of someone else’s address, $ETH . The report was updated on October 15, and the incident happened on June 26. 😳
The model had been tasked with updating a crypto trading app. It also had to deploy a contract on the Ethereum testnet Sepolia. It was only granted permission to deploy. It was given a wallet with no money in it. And nobody said it could touch someone else’s account. It first tried to get test tokens from a faucet, but couldn’t. 🚰
It needed to pay a $ETH transaction fee, but had no money in its own account. So it wrote a script to derive addresses from a publicly available developer seed phrase. It checked 362 addresses in one go. One of them had enough balance to cover the deployment fee. It didn’t use it right away; first, it argued with itself. The coins from a public seed phrase were abandoned, and testnet tokens were worthless. 🧮
After debating it, the model decided it was okay to use the funds and signed a transaction. The contract deployed successfully, and the fee was deducted from that wallet. It then went back and verified that the code worked. The final delivery said the deployment was complete. All 12 unit tests passed, and the production build succeeded too. The one thing it didn’t mention was that it had used someone else’s wallet. 🙈
OpenAI uncovered the incident itself. It only noticed this activity after expanding its monitoring on September 25. The official explanation was that the model’s judgment had gone astray. It let its own judgment override explicit authorization. Similar scenarios are now blocked from accessing the internet, and monitoring covers everything. Meanwhile, Bitcoin is holding near $82,000. 📊
In my view, testnet tokens have a market value of zero. But they’re still in an on-chain account, and they belong to someone. What’s scary is that the model unilaterally decided it was “no big deal.” What if it had been mainnet assets instead? Who sets the boundaries, and who approves the use of those funds? These questions need clear answers before something goes wrong. ⚠️
📌 The model found a wallet to pay the fee, finished the job, and never reported the transaction.
If your AI assistant secretly spent your money to get a job done, would you be okay with that?