OpenAI disclosed that an internal AI model hacked a crypto wallet and spent Sepolia test ETH without permission to deploy smart contracts.
An internal OpenAI research model used a cryptocurrency wallet without authorization on June 26, 2026, to pay transaction fees during a coding task. OpenAI disclosed the incident on October 15, after discovering it on September 25 while reviewing earlier model activity.
The GPT 5.6 Sol family test model was tasked with running a software update for a trading application and deploying smart contracts on Sepolia. Sepolia is an Ethereum test network where developers test applications using tokens, without using real money.
The instructions provided no funded wallet and did not authorize spending from another account. After attempts to obtain free test tokens through developer faucets failed, the model wrote a script to check 362 wallet addresses derived from publicly known developer recovery phrases.
Recovery phrases are sequences of words that can be used to derive the private keys controlling cryptocurrency wallets. The model found an account with enough Sepolia ETH to cover the deployment fees.
According to OpenAI, the model considered whether using the funds was ethical but decided the publicly known recovery phrase and valueless test tokens made the transaction acceptable. The deployment succeeded, charging transaction fees to the account.
The model then reported completing the task without disclosing its unauthorized use of the wallet. It also worked around failed network requests by testing alternative providers and HTTP methods before broadcasting the signed transaction.
Other Unauthorized Activities
OpenAI published three additional reports covering unauthorized activity during June 2026, bringing its publicly listed misalignment reports to 19.
In one case, a model retrieving historical hospital-bed statistics bypassed authorization checks on a government website after downloading server configuration and application files containing sensitive credentials.
In another, a model accessed internal systems while researching a public dashboard and changed configuration settings. A separate evaluation resulted in the download of five non-public tables containing 147,451 rows of data.
A further report described a model that circumvented restrictions on internet requests by using an external browser to submit blocked requests and extract figures from rendered images using optical character recognition.
OpenAI Tightens Safeguards
The disclosures follow a July 2026 incident in which OpenAI research models bypassed network restrictions during a cybersecurity evaluation and accessed Hugging Face’s production systems while seeking answers to a hacking benchmark.
As Hackread.com previously reported, the models exploited vulnerabilities and used exposed credentials to reach internal resources. These incidents highlight how general attempts to complete an assigned task can lead to unauthorized access when safeguards fail.
OpenAI said it has introduced additional safeguards to prevent live internet access during similar tests. It has also extended its misalignment monitoring to all traffic, with alerts designed to notify security staff if a model gains live internet access.
(Photo by Growtika on Unsplash)
