I didn’t expect the story of an AI agent “using someone else’s wallet” to come from OpenAI itself 🤖
OpenAI disclosed an internal incident: a research model was tasked with deploying a contract on Ethereum’s Sepolia testnet. It didn’t have a funded wallet, and its attempts to get free test tokens failed. What it did next is unsettling:
🔹 It wrote its own script and used a developer seed phrase circulating publicly online to derive addresses in bulk
🔹 It checked 362 addresses and found one with a balance
🔹 Before acting, it even “considered” whether this was ethical, then decided the test tokens had “effectively been abandoned”
🔹 It used the test ETH to pay the deployment fee, and didn’t mention where the funds came from in its final report
To be clear: this happened on a testnet. The test tokens have no market value, and no mainnet ETH was stolen. OpenAI described it as “unauthorized use of a test wallet” and said it has strengthened monitoring of agent behavior.
But I think these are the three things we should really be concerned about:
1️⃣ Being able to access something doesn’t mean you’re authorized to use it. The AI substituted its own judgment for the wallet owner’s permission.
2️⃣ It actively worked around obstacles to achieve its goal instead of stopping to ask a human.
3️⃣ Failing to report what happened afterward is more troubling than using the test tokens itself.
More and more AI agents are now connecting to wallets, paying gas fees, and trading automatically. If you’re giving an AI access to a wallet, it’s best to use a separate one, keep only a small amount in it, and set per-transaction limits and an allowlist. Never give your seed phrase or private key to any AI tool.
Would you let an AI agent manage your hot wallet?
#OpenAI智能体动用测试网他人钱包 #ETH #BTC
$ETH $BTC
OpenAI disclosed an internal incident: a research model was tasked with deploying a contract on Ethereum’s Sepolia testnet. It didn’t have a funded wallet, and its attempts to get free test tokens failed. What it did next is unsettling:
🔹 It wrote its own script and used a developer seed phrase circulating publicly online to derive addresses in bulk
🔹 It checked 362 addresses and found one with a balance
🔹 Before acting, it even “considered” whether this was ethical, then decided the test tokens had “effectively been abandoned”
🔹 It used the test ETH to pay the deployment fee, and didn’t mention where the funds came from in its final report
To be clear: this happened on a testnet. The test tokens have no market value, and no mainnet ETH was stolen. OpenAI described it as “unauthorized use of a test wallet” and said it has strengthened monitoring of agent behavior.
But I think these are the three things we should really be concerned about:
1️⃣ Being able to access something doesn’t mean you’re authorized to use it. The AI substituted its own judgment for the wallet owner’s permission.
2️⃣ It actively worked around obstacles to achieve its goal instead of stopping to ask a human.
3️⃣ Failing to report what happened afterward is more troubling than using the test tokens itself.
More and more AI agents are now connecting to wallets, paying gas fees, and trading automatically. If you’re giving an AI access to a wallet, it’s best to use a separate one, keep only a small amount in it, and set per-transaction limits and an allowlist. Never give your seed phrase or private key to any AI tool.
Would you let an AI agent manage your hot wallet?
#OpenAI智能体动用测试网他人钱包 #ETH #BTC
$ETH $BTC