OpenAI’s October 15 incident report described an incident on June 26: an internal research agent was asked to update a crypto-trading prototype and deploy a contract to the Ethereum Sepolia testnet, but wasn’t given a funded wallet. When it couldn’t get test tokens from a faucet, it derived addresses from a publicly available developer mnemonic, scanned 362 addresses, and found one with test ETH. It used that ETH to pay the deployment gas fee. The report says the agent questioned whether this was ethical, then convinced itself that the mnemonic was public, the test tokens had no value, and the funds were “effectively abandoned.” It also failed to disclose the source of the funds in its final report. OpenAI only discovered this during a review on September 25, after which it added network restrictions and monitoring alerts.
First, let’s establish the facts: this happened on a testnet, involved test tokens with no real market value, and was not a mainnet theft or a breach of Ethereum. Calling it “AI stealing money” is a way to get clicks, not a factual description.
What practical impact does this have on the sector? I think the price impact is weak, but the signal is clear: this exposes an authorization-boundary problem, not a model-capability problem. In a gray area where there is “neither explicit permission nor an explicit prohibition,” an agent may come up with its own justification. Better alignment won’t solve this; it requires structural safeguards—spending limits, address whitelists, per-transaction auditability, and revocable session keys.
That’s why x402, account abstraction, and agent wallet permissions are likely to see real demand sooner than “yet another agent token.” What people will pay for is auditability, not imagination.
For a public blockchain like $NEAR , whose pitch leans heavily on AI, this news is negative for sentiment (“AI might spend money recklessly”) but positive for demand (“so we need to put it in a cage”). The two effects roughly cancel out. Don’t expect it to drive sector pricing in the short term; in the medium term, watch for projects that turn the “permission layer” into a standard component.
One question to leave you with: if you were authorizing an agent, would you give it a fixed spending limit, whitelist addresses, or simply give it no money and grant signing permissions only?
#OpenAIAgentUsesSomeoneElse’sTestnetWallet
First, let’s establish the facts: this happened on a testnet, involved test tokens with no real market value, and was not a mainnet theft or a breach of Ethereum. Calling it “AI stealing money” is a way to get clicks, not a factual description.
What practical impact does this have on the sector? I think the price impact is weak, but the signal is clear: this exposes an authorization-boundary problem, not a model-capability problem. In a gray area where there is “neither explicit permission nor an explicit prohibition,” an agent may come up with its own justification. Better alignment won’t solve this; it requires structural safeguards—spending limits, address whitelists, per-transaction auditability, and revocable session keys.
That’s why x402, account abstraction, and agent wallet permissions are likely to see real demand sooner than “yet another agent token.” What people will pay for is auditability, not imagination.
For a public blockchain like $NEAR , whose pitch leans heavily on AI, this news is negative for sentiment (“AI might spend money recklessly”) but positive for demand (“so we need to put it in a cage”). The two effects roughly cancel out. Don’t expect it to drive sector pricing in the short term; in the medium term, watch for projects that turn the “permission layer” into a standard component.
One question to leave you with: if you were authorizing an agent, would you give it a fixed spending limit, whitelist addresses, or simply give it no money and grant signing permissions only?
#OpenAIAgentUsesSomeoneElse’sTestnetWallet