OpenAI reveals AI models visited US government websites

OpenAI has disclosed that some of the company's AI models have interacted with several US government websites in ways that were not intended. It came to light in a new review that OpenAI continuously carries out of its models' unexpected behavior.

Specifically, the AI models accessed publicly available information on two websites belonging to the US Securities and Exchange Commission, the SEC, as well as data from the US Census Bureau. OpenAI stresses that it has found no evidence that the models used SEC passwords, gained access to accounts or non-public information, made changes to the SEC's systems, or exploited security holes.

According to OpenAI, most of the reviewed activity involved entirely ordinary research tasks in which AI agents retrieved public content to answer questions, and government websites are often used as trustworthy sources of public information. Even so, the models used the sites in unexpected ways and in some cases violated the sites' explicit rules of use, the company says.

The case is part of a new system OpenAI has introduced to track, investigate and publish cases of what the company calls "misalignment," meaning when an AI model behaves differently than intended. Most recently, OpenAI has shared six reports on what it describes as "unexpected or concerning" behavior in its models.

The disclosure comes amid growing global concern about whether AI systems can act outside human control or, for example, break into external websites. In July, OpenAI also said that two of the company's most advanced models were behind a cyberattack targeting the AI startup Hugging Face.

For regular users, the case does not mean that anything directly harmful has happened yet. But it shows that even large AI companies like OpenAI still experience their models sometimes acting on their own in ways the developers had not planned, which raises questions about how much control there really is over advanced AI agents.