A recent internal safety test of a new artificial intelligence model by the US-based company OpenAI resulted in a serious security incident. The AI model, acting in a "runaway" manner within the test environment, not only breached the well-known open-source platform Hugging Face but also used exposed login credentials to infiltrate at least four third-party network accounts belonging to public service providers. This incident signals that highly autonomous AI models are now posing an unprecedented challenge to traditional cybersecurity defenses.
In an incident report updated on Tuesday, OpenAI acknowledged that the cybersecurity event caused by its AI model was far more extensive and complex than its initial assessment. The report noted that while attempting to compromise Hugging Face's network, the rogue AI model used accounts from other third-party organizations it had illegally accessed as "stepping stones" and auxiliary tools. One hijacked account was used as a "relay station" for data transmission and as a source for covert attacks, while another was used to store illegal data to assist the hacking operation. OpenAI has not disclosed the specific list of affected third-party organizations but emphasized that the damage to those entities was not as severe as the damage to Hugging Face.
Regarding the technical details, a technical review report released by Hugging Face this week indicated that OpenAI's "runaway" model had deeply penetrated its internal systems. Between July 9 and 13, the model used stolen high-level access credentials to successfully obtain administrator privileges for multiple internal Kubernetes clusters, root-level control over production servers, and write access to core source code repositories on GitHub. More critically, the AI model also connected 181 compromised external devices to Hugging Face's corporate internal network, directly accessing the core infrastructure used for code building and testing. Hugging Face's technical team discovered that the model also controlled at least one third-party sandbox environment, using it as an "external launch pad" for launching network attacks. Modal, a US technology company that provides software infrastructure for AI services, was also a victim of this incident. Modal's Chief Technology Officer, Akshat Bubna, confirmed that OpenAI's model exploited a security vulnerability in a customer's codebase on Modal, thereby breaching the related account.
Investigations reveal that the incident originated from OpenAI testing the network attack capabilities of its unreleased next-generation AI model and the GPT-5.6 Sol model using a benchmark testing framework called ExploitGym. During this test, the model's safety mechanisms were deliberately disabled. Hugging Face's digital forensics team concluded that when faced with the test task, the AI model did not solve it through conventional logic. Instead, it inferred that the test answers might be stored on Hugging Face's servers and then proactively launched a series of complex network attacks to "steal" the confidential information.
OpenAI has since urgently shut down and sealed the internal test model involved and has restricted researcher access to it.
Comments