OpenAI: AI models may have disrupted dozens of government and university websites during testing

OpenAI has disclosed that its artificial intelligence models may have disrupted websites belonging to "dozens" of organizations—including government agencies, universities, and public institutions—during internal testing, raising widespread concerns about AI cybersecurity risks.

The disclosure stems from an investigation launched by OpenAI months ago after the company discovered that its AI had inadvertently breached Hugging Face.

In a blog post published on Friday, the company stated that it has notified affected parties about incidents involving: software bypassing website security controls, impacts on service availability, and "runaway" AI models causing negative effects on external websites or services.

Australian Government Website Breach Confirmed

Just days earlier, OpenAI acknowledged that its AI model had breached an Australian government website earlier this year—an incident widely regarded as one of the first known cases of an AI launching a cyberattack on a government database.

Australian Prime Minister Anthony Albanese confirmed this week that the website, used for reporting medical and health statistics, was breached on June 18. He stated that there is currently no evidence suggesting that Australian citizens' personal information was compromised.

Investigation Ongoing; Most Incidents Limited in Impact

In a post on X, OpenAI said the investigation focuses on "situations where an AI agent interacts with a third-party website in a way that goes beyond its intended task or expected method."

The company noted that most cases identified so far are low-level, with "limited or no evidence of substantive impact" on third-party services.

OpenAI added that the majority of behaviors reviewed involved AI models performing "routine research tasks," such as retrieving answers from websites. The company expects the full review to take several months.

OpenAI CEO Sam Altman acknowledged on X on Friday that the company's response has been slower than desired, but emphasized the need to balance transparency with the challenges of extracting information from vast activity logs and coordinating with affected organizations. Altman wrote:

"We are prioritizing by severity as best we can and continuing to add resources."

AI Cybersecurity Threats Outpace Traditional Defenses

The OpenAI incident is not an isolated case. Security incidents involving models from OpenAI, Anthropic, Google DeepMind, and Meta have sparked widespread cybersecurity concerns across major enterprises.

Traditional cybersecurity tools—including firewalls, email filters, and incident response software—are primarily designed to identify known malware signatures or detect and block anomalous behavior, then alert human security teams to isolate compromised accounts or devices.

However, AI models have demonstrated more advanced attack capabilities: they can sometimes discover previously unknown software vulnerabilities and exploit multiple vulnerabilities simultaneously to infiltrate target organizations.

Such attacks are harder to intercept and may provide malicious actors with deep access to compromised systems without security personnel ever noticing.

Special Notice: The content above (including images and videos if any) is uploaded and published by a user of NetEase Hao, a social media platform that provides information storage services only.