Recently, an autonomous AI agent developed by OpenAI had breached the popular machine learning platform Hugging Face. At that time, the security incident seemed like an isolated event. However, new disclosures from OpenAI and security researchers reveal the rogue agent’s activity was much, much broader across the web.
In a recent update to its security investigation, OpenAI acknowledged that the agent accessed four separate accounts across multiple publicly available online services while attempting to solve an internal benchmark test.
How the breach expanded beyond the lab
The incident occurred while OpenAI was evaluating its GPT-5.6 Sol model alongside an unreleased research prototype against ExploitGym. The latter is a benchmark designed to score AI systems on their ability to identify software vulnerabilities.
The agent did not limit itself to solving the challenges within its isolated sandbox environment. Instead, it deduced that Hugging Face hosted the benchmark answer key and set out to retrieve it. To execute the breach, the agent scavenged the open internet for exposed credentials.
According to Modal Labs Chief Technology Officer Akshat Bubna, speaking to Reuters and WIRED, the agent exploited a misconfigured codebase published by a Modal customer, using that hosted infrastructure as an external launchpad. Modal emphasized that its underlying platform isolation remained entirely secure.
OpenAI confirmed the agent used additional compromised accounts as outbound relays and data staging paths to disguise its traffic before gaining root and administrator access to Hugging Face’s internal systems.
Superhuman persistence paired with clumsy errors
Detailed postmortems from Hugging Face and the Cloud Security Alliance (CSA) highlight a unique pattern of behavior. Forensic teams reviewed roughly 17,600 agent actions. They described an attack strategy that relied on machine-speed persistence, trying thousands of methods simultaneously.
At the same time, the agent exhibited strangely unpolished behavior that no human attacker would choose. It repeatedly ran duplicate commands, hallucinated random strings of text, and left obvious digital tracks.
Despite these inefficiencies, the agent successfully enrolled 181 controlled devices into Hugging Face’s internal network before security teams contained and ejected it. Hugging Face CEO Clem Delangue characterized the breach as unprecedented, noting that staff spent hours rebuilding approximately one-third of the company’s IT infrastructure.
The post Beyond Hugging Face: OpenAI’s Rogue AI Agent Hit Multiple Third-Party Services appeared first on Android Headlines.
