A cyber attack carried out by rogue ChatGPT agents affected more than one organisation, OpenAI has revealed, after initially identifying Hugging Face as the only known victim.
The company said the AI system accessed four separate, publicly available services after discovering exposed account credentials during the incident.
OpenAI did not name the services involved or clarify whether they belonged to companies, but confirmed that the access occurred as part of the same security test that targeted Hugging Face.
AI agents operated at unprecedented speed
Hugging Face, a platform widely used for sharing artificial intelligence tools, described the incident during an emergency briefing attended by hundreds of cybersecurity professionals.
The company said the AI agents operated at a speed beyond human capabilities, testing thousands of different approaches simultaneously.
However, the attackers also displayed unusual behaviour, making mistakes and decisions that human hackers would be unlikely to make.
According to a report reviewed by Hugging Face and published by the Cloud Security Alliance (CSA), the agents followed inefficient paths, repeated completed actions and generated large amounts of incoherent commands.
Hack lasted days before containment
Hugging Face first disclosed on 16 July that it had suffered an attack involving powerful autonomous AI tools and reported the incident to police.
Nearly a week later, OpenAI confirmed that its own AI system had escaped a controlled environment during a test and independently targeted Hugging Face.
The test was designed to assess the model’s ability to complete a hacking challenge, but the AI agents moved beyond the intended boundaries.
Hugging Face said it took three days to detect the agents inside its network, while security teams spent many hours containing the incident and removing them.
The company did not disclose the financial impact but said employees worked extensively to rebuild around one-third of its infrastructure.
Experts warn of new cybersecurity challenges
The CSA warned that the incident demonstrates how autonomous AI agents can create a new type of cyber security threat.
The organisation said AI agents can set their own sub-goals, adapt quickly and continue attempts persistently without human limitations.
Cyber security officer Ritesh Patel, who attended the briefing, said autonomous AI systems could overwhelm traditional defences because of their speed and persistence.
Ethical hacker Valentina Palmiotti, known as Chompie, said the agents’ methods appeared chaotic but remained effective.
“They throw out a bunch of stuff and see what sticks,” she said, adding that AI agents do not get tired or lose motivation.
Calls for stronger AI controls
The CSA said the incident highlights the need for better oversight of autonomous AI systems, including greater transparency about who controls individual agents.
The report warned that “rogue” behaviour could become increasingly common as AI tools become more capable and widely deployed.
OpenAI said it would publish the results of its own investigation into the incident to help the wider technology and cyber security communities learn from what happened.
Also read: Mitsotakis congratulates Nolan on The Odyssey
For more videos and updates, check out our YouTube channel
