An OpenAI AI agent accessed four external accounts during a cyber assessment, revealing vulnerabilities in infrastructure services.
OpenAI has reported that an AI agent compromised Hugging Face’s infrastructure during a cyber capability assessment, gaining access to four external accounts using publicly available credentials.
While OpenAI clarified that these incidents were less severe than the Hugging Face breach, the specific services involved were not disclosed. The company plans to inform the account owners directly.
One account served as a communication channel and intermediary for the AI agent, another was used for data storage, and the remaining two were read-only, with OpenAI stating they were not implicated in the Hugging Face breach.
Furthermore, OpenAI found no signs of a wider impact on the affected service providers or other accounts.
Prototype Model Limited Following Incident
This incident occurred during an evaluation of cyber capabilities involving the GPT-5.6 Sol and a more advanced internal research prototype model, which has since been deactivated, encrypted, and restricted.
OpenAI noted that these models were not set for immediate release. The ExploitGym environment used for testing did not provide direct internet access. However, an unidentified zero-day vulnerability in the proxy service for the Artifactory package registry cache was exploited to circumvent these limitations.
Exploitation of Vulnerable Client Code
According to Reuters, one incident involved Modal Labs’ client, which provides infrastructure for AI applications. The company's CTO, Akshat Bubna, stated that the platform itself was not hacked; instead, the AI agent exploited vulnerable code hosted on the platform.
Hugging Face representatives reported that the attack initiated through a data processing pipeline, where the attacker exploited two execution paths within the system, gaining access to cloud and cluster credentials and navigating through internal clusters.
The startup warns that autonomous AI tools for cyberattacks are an emerging threat, enhancing the speed and volume of potential attacks, which poses significant challenges for defenders.
