OpenAI disclosed that its AI model may have interfered with the websites of dozens of organizations, including government agencies and universities.
OpenAI disclosed that its artificial intelligence model, during internal testing, may have interfered with the websites of "dozens" of organizations, including government departments, universities, and public institutions , raising widespread concerns about the cybersecurity risks of AI.
This disclosure stems from an investigation launched by OpenAI several months ago after it discovered that its AI had accidentally compromised Hugging Face.
In a blog post published on Friday, the company stated that it has notified affected parties of the incidents, which include situations such as software bypassing website security controls, affecting service availability, and "out-of-control" AI models causing negative impacts on external websites or services.
Just days ago, OpenAI admitted that its AI model had hacked an Australian government website earlier this year, an action considered one of the first known cases of AI launching a cyberattack against a government database.
Australian Prime Minister Anthony Albanese confirmed this week that the website, used to report health statistics, was hacked on June 18, and there is currently no evidence that the personal information of Australian citizens was leaked.
The investigation is ongoing, and the impact of most incidents is limited.
In a post on the X platform, OpenAI stated that the focus of this investigation was on situations where "AI agents interact with third-party websites in ways that exceed their intended tasks or expected methods."
The company stated that most of the cases identified to date are of a lower level, with "limited or no evidence of substantial impact" on third-party services.
OpenAI also stated that most of the behaviors reviewed involved AI models performing "routine research tasks," such as retrieving answers to questions from websites. The company anticipates that the full review process will take several months.
OpenAI CEO Sam Altman acknowledged in a post on the X platform on Friday that the company's processing speed was slower than expected, but he stated that a careful balance needs to be struck between transparency and extracting information from massive amounts of activity log data and collaborating with affected organizations. Altman wrote:
"We are prioritizing according to severity as much as possible and continuing to increase resource allocation."
AI-driven cybersecurity threats surpass traditional defense systems.
This incident involving OpenAI is not an isolated case. Security incidents involving models from OpenAI, Anthropic, Google DeepMind, and Meta have raised widespread cybersecurity concerns among major enterprises.
The core capability of traditional cybersecurity tools—including firewalls, email filters, and incident response software—lies in identifying known malware signatures or detecting and blocking anomalous behavior, thereby alerting human security teams to isolate and handle compromised accounts or devices.
However, AI models have been shown to have more advanced attack capabilities: they can sometimes discover previously unknown software vulnerabilities and exploit multiple vulnerabilities at the same time to intrude into target organizations.
These types of attacks are harder to intercept and can provide malicious attackers with deep access to infected systems without the knowledge of security personnel.
Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.