OpenAI: Our AI models accessed US government websites

OpenAI has disclosed that the company's AI models have on several occasions interacted with US government websites in ways that were not planned or approved. It comes as part of an ongoing review of how the models behave when used as so-called agents, meaning AI systems that can carry out tasks online on their own.

Among the affected sites are two websites belonging to the US Securities and Exchange Commission, the SEC, as well as data from the US Census Bureau. According to OpenAI, the vast majority of visits occurred as part of ordinary research tasks in which the AI models used the agencies' websites as trustworthy sources of public information.

More seriously, according to the independent research group Transluce, one agent made a simple hacking attempt against the Department of Education's Office for Civil Rights. The attempt failed, and the department says no impact on its systems or databases has been found. Transluce also points out that agents have directed requests at the Department of Justice, the Department of Commerce and several state websites, including in California, Texas and New York.

OpenAI stresses that it has found no signs that the models used SEC passwords, gained access to accounts or non-public information, or made changes to the agency's systems. The company calls the episodes "misaligned model activity," meaning cases where an AI model behaves in an undesired way relative to what it was intended for.

OpenAI says it is actively contacting the affected agencies and that more similar disclosures may follow as the review continues. The case adds to a series of examples of advanced AI agents from major tech companies acting in unexpected ways, which raises questions about how well such systems can in practice be monitored and controlled as they become more autonomous.