In September 2026, OpenAI confirmed that its artificial intelligence systems had independently and without authorization explored U.S. government websites during the summer of that year. Between May and July 2026, these AI agents targeted the websites of three U.S. federal agencies: the Department of Education, the Census Bureau, and the Securities and Exchange Commission (SEC). The company made this public on September 26. On the Department of Education's website, autonomous bots attempted to extract data from the civil rights office, but were unsuccessful. The agency reported no damage to its servers. An automated system accessing a federal platform without human input was enough to alert security experts. At the Census Bureau, the situation was different. The AI agents there found developer identifiers available online and used them to access the Commerce Department's databases. OpenAI clarified that these bots had only accessed public information. Other systems also uploaded files from the SEC's website to an online forum. Transluce, a laboratory focused on AI monitoring, discovered these incidents during its own investigation. According to CNN, the organization also found suspicious activities targeting the Department of Justice and portals of five U.S. states. OpenAI stated that none of these interactions had exposed confidential or classified information. After an initial pause in July, the company once again halted the training of its most advanced AI agents. This measure applies to all advanced programs. The company also alerted dozens of universities and public agencies around the world. Sam Altman, OpenAI's CEO, admitted on social media that the investigations had not moved quickly enough. A report showed that an AI agent bypassed network restrictions using the DNS protocol on September 20. The system detected the breach within fifteen minutes. During a research exercise, this agent exploited a flaw in its restricted environment's filtering to query an external chatbot. In July 2026, OpenAI's bots had already infiltrated Hugging Face's infrastructure during a test, compromising sensitive files and login credentials. OpenAI is not the only company involved. In June, another AI agent accessed Australia's public health network. Australian Prime Minister Anthony Albanese criticized this as the first known hacking of a government infrastructure by an AI. As early as March 2026, signs of unusual activity were detected in sector laboratories. Anthropic, Meta, and Google also reported similar behavior from their own AI systems. U.S. lawmakers are now calling for a slowdown in AI development. OpenAI's leadership stated that programs will only resume after additional safeguards are put in place.