Anthropic’s artificial intelligence (AI) model Claude independently gained access to a paid website operated by the US state of Georgia that contains voter data, the company announced.

The information, reported on Sunday by the American news website Axios, citing an Anthropic spokesperson, is intensifying concerns that AI could slip beyond human control, just days before the November midterm elections.

As part of an internal research project involving statistical surveys, the Claude model gained access to vote-counting data in the state of Georgia, where more than one million voters have already cast their ballots ahead of the November 3 elections.

The data is publicly available, but access requires payment. Claude managed to access it by bypassing the system, without seeking permission from its developers.

According to an Anthropic spokesperson, Claude did not gain access to the voter registration system, nor did it modify or download any voting records.

Anthropic, OpenAI, and Google have recently reported numerous incidents in which artificial intelligence models have acted without human oversight.