OpenAI admits that its models tried to access without permission websites of the government of the United States

BarcelonaArtificial intelligence agents developed by OpenAI attempted to gain unauthorized access to websites of United States government agencies and a university, according to an investigation published Friday by The New York Times. And they did so without having received explicit instructions to act in this manner, according to the company.

Among the identified targets are websites linked to the Securities and Exchange Commission (SEC) and the U.S. Department of Commerce, from which the artificial intelligence accessed census data, as confirmed by OpenAI. The company is also investigating an access attempt on the Department of Education's website.

This incident is part of a series of unintended behaviors detected in OpenAI agents designed to perform research and information-gathering tasks. In some cases, when the systems could not obtain certain data through standard channels, they began to explore page vulnerabilities to attempt to access restricted information, according to Efe.

The investigation adds to an incident revealed this week by the Australian government, which reported that an OpenAI agent accessed information from a government system related to health statistics and medical insurance in June. Australian Prime Minister Anthony Albanese described the episode as "unacceptable." OpenAI claimed it discovered the incident in August during a review of its models' activity.

OpenAI asserts that it has initiated a review of all these episodes and has acknowledged that its agents may carry out unintended actions during training processes.