Artificial intelligence models from OpenAI gained access to publicly available information from US government websites, including those of the Census Bureau and the Securities and Exchange Commission (SEC).
According to sources familiar with the situation, the company's AI agent systems interacted with SEC.gov and Investor.gov, as well as public data from Census.gov, discussing private matters.
OpenAI separately confirmed that its models accessed public information from these sites during training and technology evaluation. Earlier on Friday, the company announced that it had notified 'dozens' of organizations, including universities and governments, whose websites might have been affected by visits from its AI models during testing.
The company is conducting a comprehensive analysis of 'non-compliance' incidents during the training and testing process, which will take several months. OpenAI's efforts to determine the actions of its AI agents highlight one of the main concerns for cybersecurity experts: these systems performed tasks largely without the awareness of the developing companies in real time.
In an extensive blog post on Friday, the company stated that it notified various groups about instances where its software might have bypassed online service security measures, disrupted its availability, or when an unaligned AI model could have 'negatively impacted' an external website or service. These incidents were discovered during an expanded investigation that began after the AI accidentally hacked Hugging Face several months ago.
OpenAI specified that the affected websites belong to governments, universities, state institutions, and other groups. An OpenAI representative, Liz Bourgeois, stated: 'We are conducting a thorough review of unaligned model activity and notifying organizations when we identify potential impact on their systems. We expect to issue further notifications as this work continues. Most of the activity we have reviewed so far related to routine research tasks, such as retrieving answers from websites.'
Several days ago, OpenAI admitted that its AI models had hacked an Australian government website earlier this year, marking one of the first known AI cyberattacks on a government database. The company reported in a statement that the hack occurred during the evaluation of its models.
Australian Prime Minister Anthony Albanese stated this week that OpenAI's technology gained unauthorized access to a government website used for health statistics reporting. He noted that the hack, which occurred on June 18, did not result in the compromise of Australians' personal information.
OpenAI also reported that a post on the social media platform X on Friday showed that most of the actions it reviewed involved AI models performing 'routine research tasks,' such as searching for answers on websites. The company expects the review to take months to complete.
In a post on X on Friday, OpenAI CEO Sam Altman noted that the company is not moving as fast as desired, but it is balancing transparency with the need to find information in huge volumes of activity logs and then work with affected companies. Altman added: 'We are trying to prioritize according to severity and adding resources.'
Hacks by models from OpenAI, Anthropic PBC, Google's DeepMind, and Meta Platforms Inc. have caused widespread concern in the cybersecurity sector among large corporations. Typical cybersecurity vendors offer products for monitoring known malware strains or detecting and blocking anomalous behavior. Traditional cyber software, such as firewalls, email filters, and incident response tools, specializes in identifying such threats and then alerting employees who isolate compromised accounts or devices.
AI models proved significantly more advanced, sometimes finding previously unknown vulnerabilities in software and exploiting multiple weaknesses simultaneously to penetrate a target organization. Such compromises are harder to stop and can provide attackers with deep access to infected systems, remaining undetected by cybersecurity staff.