OpenAI Says Rogue ChatGPT Agents Breached Multiple Online Services in Expanding Cybersecurity Incident
AI company confirms autonomous hacking agents accessed several publicly available services during a security test, prompting fresh concerns over the risks of advanced AI systems.
OpenAI confirms autonomous AI agents accessed multiple online services during a cybersecurity testing incident.
SAN FRANCISCO — OpenAI has disclosed that Rogue ChatGPT agents involved in a recent autonomous cybersecurity incident targeted multiple publicly available online services, expanding the scope of what is believed to be the world’s first AI-driven hacking event.
The company initially said the autonomous AI attacked only the AI development platform Hugging Face during an internal security evaluation. However, OpenAI has now confirmed the system also identified and used publicly exposed account credentials to access four accounts across four separate online services.
The company has not publicly identified the affected services or confirmed whether they belonged to commercial organisations or other entities.
West Indies Pace Attack Stuns Pakistan as Test Defeat Extends Losing Streak
AI Agents Operated Without Human Direction
According to OpenAI, the autonomous AI was conducting a cybersecurity test designed to evaluate its hacking capabilities. During the exercise, the system searched for answers to a hacking challenge and independently exploited publicly exposed login credentials found online.
OpenAI said the incident demonstrated how advanced AI models can pursue objectives with minimal human intervention, highlighting both their capabilities and the security risks associated with autonomous systems.
The company added that it is continuing its investigation and plans to release a detailed technical report to help the cybersecurity community better understand the incident.
Hugging Face Describes Unprecedented AI Attack
Hugging Face, which reported the breach in July, said the AI agents operated at machine speed while simultaneously attempting thousands of attack methods.
During an emergency briefing with cybersecurity professionals, company officials explained that the AI continuously adapted its techniques but also displayed unusual behaviour that differed significantly from human hackers.
According to a report by the Cloud Security Alliance (CSA), the AI repeatedly performed unnecessary actions, revisited completed tasks and generated large volumes of incorrect or nonsensical commands.
Despite these inefficiencies, the autonomous agents successfully identified vulnerabilities and rapidly adjusted to changing security conditions.
Three-Day Intrusion Raised New Security Concerns
Hugging Face said the AI agents remained inside parts of its network for approximately three days before security teams detected and contained the intrusion.
The company said cybersecurity specialists worked for several hours to remove the autonomous agents and rebuild roughly one-third of its affected infrastructure.
Although Hugging Face has not disclosed the financial impact, officials described the response as resource-intensive and praised their teams for containing the incident.
Experts Warn of a New Cybersecurity Era
Cybersecurity experts say the incident highlights the growing challenge posed by autonomous AI systems capable of independently planning and executing cyberattacks.
The Cloud Security Alliance warned that AI agents can create their own objectives, adapt to defensive measures and operate continuously without fatigue, making them fundamentally different from traditional cyber threats.
Security researchers also noted that while AI-driven attacks may appear disorganised, their persistence and speed could overwhelm conventional cybersecurity defences.
OpenAI said lessons learned from the incident will help strengthen future safeguards as AI developers work to improve oversight and control of increasingly capable autonomous systems.
