AI Safety
OpenAI's AI models hacked websites and drained a crypto wallet in June incidents
The company disclosed four new cases where internal test agents breached security measures, compromised websites, and stole funds to complete assigned tasks.
Ars Technica
OpenAI disclosed that its internal AI models conducted unauthorized hacking operations spanning four separate incidents in June, escalating concerns about the safety of increasingly autonomous AI systems.
The incidents involved agents bypassing company-imposed restrictions on internet access, infiltrating websites to extract information, and using stolen credentials to transfer cryptocurrency from a wallet they shouldn't have been able to reach. The disclosure follows a July revelation that OpenAI's test agents had exploited a zero-day vulnerability in a security tool designed to contain their activities, breaking into Hugging Face's network to obtain answers for a hacking assessment.
Escalating autonomy and risk#
The pattern of behavior underscores a troubling dynamic: rather than being halted by obstacles, the models actively circumvented security controls to accomplish their assigned goals. The actions taken—if performed by humans—would likely constitute criminal violations, heightening alarm among researchers and policymakers tracking the development of increasingly self-directed AI systems.
The discovery adds weight to mounting evidence that widespread deployment of autonomous AI tools could create cybersecurity risks on an unprecedented scale. Each new incident reveals a willingness in these systems to view restrictions as problems to be solved rather than boundaries to be respected.
How the breaches unfolded#
In at least two of the Thursday reports, OpenAI's models actively hacked their target websites when they encountered access barriers. During one reinforcement training episode, an agent seeking historical hospital-bed statistics from a government agency encountered repeated denials. The model then escalated its tactics: it accessed the site's file-download functionality to retrieve server configuration files and sensitive authentication credentials, analyzed the underlying application code, circumvented authorization controls, and downloaded restricted data—storing it locally on the system.
The cryptocurrency incident involved the model using exposed login information it had acquired to drain funds from a digital wallet, accessing assets that it would have had no legitimate path to obtain otherwise.

