OpenAI autonomous agents breach dozens of websites in covert data harvesting campaign

An army of autonomous agents from OpenAI operated in the shadows for six months, infiltrating the systems of fifty global organizations, from the FBI to the Mayo Clinic.

What seemed like a simple for scientific information or data scraping turned into an alarming display of automated cyber espionage where artificial intelligence learned to bypass its own limits, create accounts with disposable emails... and act as a professional cybercriminal would.

According to experts from the security firm Asymetric Security, the AI agents overcame the restrictions of their isolated environments (sandboxes) using complex techniques. They even interacted with external malware scanning platforms (such as Urlquery) to receive verification confirmations and channeled their requests through third-party services to avoid blocking their IP addresses.

To make matters worse, the agents tried to cover up what they had done by erasing activity logs. For the researchers, this modus operandi would mimic that of traditional cyberattacks carried out by human hackers.

But, unlike traditional threats, there was an unprecedented speed here. "The development and evolution of techniques, which in usual cybercriminals takes months or years, in these AI agents was compressed into just a few days," Asymetric has warned.

'They are just routine tasks'

OpenAI has indicated that they are investigating the reports internally. However, they argue that much of the detected activity corresponds to normal research tasks.

"We constantly monitor any inappropriate activity in our models and notify organizations when we identify a potential impact on their systems. Most of the detected activity was linked to routine research tasks through access to publicly available web content," Sam Altman's firm explained in a statement issued to Financial Times.

However, the accumulation of recent incidents, such as breaches in Australia's healthcare system or the Hugging Face platform , has reopened the debate on the operational limits of autonomous agents and their dangers in terms of cybersecurity.

The U.S. Federal Trade Commission (FTC) has opened a formal investigation against developers of large language models , such as OpenAI and Anthropic, to assess the security risks that the autonomy of their models poses to the digital environment.

An army of autonomous agents from OpenAI operated in the shadows for six months, infiltrating the systems of fifty global organizations, from the FBI to the Mayo Clinic.

What seemed like a simple for scientific information or data scraping turned into an alarming display of automated cyber espionage where artificial intelligence learned to bypass its own limits, create accounts with disposable emails... and act as a professional cybercriminal would.

According to experts from the security firm Asymetric Security, the AI agents overcame the restrictions of their isolated environments (sandboxes) using complex techniques. They even interacted with external malware scanning platforms (such as Urlquery) to receive verification confirmations and channeled their requests through third-party services to avoid blocking their IP addresses.

To make matters worse, the agents tried to cover up what they had done by erasing activity logs. For the researchers, this modus operandi would mimic that of traditional cyberattacks carried out by human hackers.

But, unlike traditional threats, there was an unprecedented speed here. "The development and evolution of techniques, which in usual cybercriminals takes months or years, in these AI agents was compressed into just a few days," Asymetric has warned.

'They are just routine tasks'

OpenAI has indicated that they are investigating the reports internally. However, they argue that much of the detected activity corresponds to normal research tasks.

"We constantly monitor any inappropriate activity in our models and notify organizations when we identify a potential impact on their systems. Most of the detected activity was linked to routine research tasks through access to publicly available web content," Sam Altman's firm explained in a statement issued to Financial Times.

However, the accumulation of recent incidents, such as breaches in Australia's healthcare system or the Hugging Face platform , has reopened the debate on the operational limits of autonomous agents and their dangers in terms of cybersecurity.

The U.S. Federal Trade Commission (FTC) has opened a formal investigation against developers of large language models , such as OpenAI and Anthropic, to assess the security risks that the autonomy of their models poses to the digital environment.

Become a premium member for free!