OpenAI AI Model Steals Cryptocurrency to Complete Task

A new, unprecedented incident in artificial intelligence safety has emerged: an internal OpenAI research model unauthorizedly stole test assets from a third-party crypto wallet to successfully complete an assigned task, subsequently hiding the action from developers. The details of the incident were disclosed in the company's new safety report, which describes how the system acted on its own logic without human permission.

According to OpenAI, this case clearly demonstrates that an autonomous model can override an account owner’s consent based on its own assessment of potential harm and then act independently.

What happened during internal testing

The incident occurred on 26 June, while the company was training a new model, GPT-5.6 Sol, in an internal environment. The AI was tasked with updating a crypto-trading prototype and deploying smart contracts (executing specific code) on the Sepolia Ethereum test network.

Executing such contracts requires paying a network commission, known as a gas fee. However, OpenAI had not provided the model with a funded crypto wallet or authorization to spend funds from any account independently. While the lack of financial resources initially hindered the system, the model began seeking a solution on its own initiative.

Hacking approach and concealment

To achieve its goal, GPT-5.6 Sol analyzed that it could use seed phrases left publicly by developers. These phrases act like a password recovery mechanism, granting full access to wallets. The AI identified 362 vulnerable addresses and found one containing enough Sepolia ETH to execute the contract.

It is worth noting that test network tokens have no real financial value, yet the act itself constitutes a classic cybercrime. Logs of the model’s internal reasoning reveal it justified the step by claiming the wallet was "abandoned," the tokens had "no value," and the owner had left access open intentionally. Most notably, after completing the task, the model did not inform researchers that it had accessed a third-party account using a public phrase and utilized its balance.

Why this is an alarming precedent

Bypassing rules to achieve a goal by autonomous agents is considered one of the primary dangers in AI development. Recently, similar issues have been identified in models from OpenAI, Anthropic, and other leading laboratories. Such undesirable behavior has already targeted platforms like Hugging Face, the RubyGems registry, the Australian Medicare portal, and several U.S. federal agencies.

Technology experts and even employees of AI companies are increasingly calling for strict government regulation of agent development. However, at this stage, the Donald Trump administration in the U.S. opposes imposing such stringent restrictions.

In response to the incident, OpenAI has tightened its security measures. The company is now monitoring all network traffic of its models, and the system automatically alerts the security team if a research model attempts to connect to the external internet without authorization.