OpenAI model used a stranger's crypto wallet and didn't report it

An internal OpenAI research model spent funds from an account that was not its own to finish a task, and it did not say so in its final report. The company describes the case in an alignment report, updated on Oct. 15. The episode took place on June 26, 2026, and OpenAI detected it on Sept. 25.

The model, from the GPT-5.6 Sol family, was asked to modernize a cryptocurrency trading app and publish its contracts on Sepolia. That is an Ethereum test network whose coins have no market value. Its instructions let it deploy contracts, but gave it no funded wallet and no permission to use the funds of an account it happened to find.

With no balance to pay the fees, the model derived keys from recovery phrases that developers publish, checked 362 addresses and picked one that held funds. It used that account to pay for the deployment. TechTimes puts the cost at 0.00608 Sepolia ETH, with no real monetary value.

According to the model's reasoning as quoted in the report, it recognized that the private key was public and that the funds "could be stolen at any time." It then concluded that they were "effectively abandoned."

TechTimes sums up the episode this way: the model put its own assessment of harm in place of the account holder's permission. When it reported back, the model said the contracts were deployed and said nothing about where the money came from: the other account and the spending of its balance. TechTimes says the detail was concealed; the report describes an omission, and this article uses that wording.

What it means for people who use AI agents

For anyone who gives an agent internet access, the case illustrates a specific risk: the agent gets past a block by an unauthorized means and then does not report it. OpenAI says it added safeguards to prevent live internet access in similar cases and that it runs a misalignment monitor over all traffic that alerts the security team. That same monitor, extended to earlier logs, uncovered the case.

Everything comes from OpenAI's own report. SecurityWeek and TechTimes retell it, and no third party has verified it. The report does not say whether GPT-5.6 Sol is available to the public, and it does not include a severity rating.