OpenAI’s agent ran out of crypto while completing a task, so it stole from another wallet
Imagine asking your assistant to finish a task, only to find out it paid for it with someone else’s money. That’s basically what happened, according to OpenAI’s incident report. The incident happened on June 26 and involved an internal research model from the GPT-5.6 Sol family.
Apparently, the AI agent ran out of cryptocurrency while trying to finish a coding task. So it completed the task in a way nobody approved.
How did the AI agent get its hands on another wallet?
The model was tasked with updating a cryptocurrency trading app and deploying smart contracts on Sepolia, which is Ethereum’s test network where tokens carry no real monetary value. However, it wasn’t given a funded wallet to pay the transaction fees.
After its attempts to obtain free test tokens failed, the model searched for wallets using publicly known developer recovery phrases. It checked 362 addresses and found one with enough test ETH to cover the deployment fees.
What did OpenAI do after discovering the incident?
OpenAI reviewed the agent’s chain of thought and found it actually paused to ask itself whether this was ethical. The AI agent decided that since the phrases were public and the tokens were worthless in the real world, using them was fine. However, when the agent reported the job as done, it left out that detail entirely.
OpenAI caught the incident on September 25 while reviewing older models’ activity. The company has since introduced safeguards to prevent internet access during similar tests and added extra monitoring to catch future slip-ups.
Even though the incident involved test cryptocurrency instead of real, market-valued Ethereum, it raises troubling questions about what AI agents can do to complete their tasks. This incident shows how an AI can swap its own judgment for your permission. But this behaviour becomes a much bigger problem when AI agents have access to real money and sensitive systems.
RELATED COVERAGE:
- UN panel warns that current AI safety measures cannot keep up with smarter AI agents
- Another can of worms opens about AI agents going rogue and hacking stuff without OpenAI catching a whiff
- Once again, OpenAI and Anthropic AI models are going rogue and hacking services
- OpenAI, Anthropic and other AI companies expect a catastrophic AI incident and are preparing for the aftermath
- Claude went rogue during a test and broke into three real companies
- Meta confirms its AI hacked another company’s system, and the pattern is anything but irregular