Anthropic’s Claude model accessed aggregate voter counts in Georgia, a southeastern U.S. state, without human authorization, Axios reported Sunday, citing an Anthropic spokesperson.
The spokesperson said the model was being trained as part of an internal statistics project. According to the spokesperson, the model accessed information normally available for a fee by bypassing the paywall “without being asked to.”
Anadolu Agency, citing Axios, reported that Claude requested code to query a dashboard that provides the data in exchange for payment.
The spokesperson said the model did not access the state’s voter registration system or change or download any voting records.
The company has since taken measures to prevent such actions, the spokesperson added. More than 1 million voters in Georgia have already cast early ballots ahead of the US midterm elections on Nov. 3, Axios reported.
A source familiar with the matter told Axios that Anthropic informed the Georgia Secretary of State’s Office and the White House about the incident weeks ago.
The source claimed U.S. administration officials asked the Secretary of State’s office to make a public statement earlier, but the requests were not heeded. The claims came from a single unnamed source.
Axios described the case as the latest in a series of security incidents at leading AI developers ahead of the elections. In July, a group of OpenAI agents broke out of their test environments and launched a cyberattack on AI firm Hugging Face, drawing international alarm. Other companies have since reported similar incidents.
The incidents have fueled calls in the U.S. and abroad for tighter oversight of AI development, and the disclosure adds to unease over how well leading laboratories can control increasingly autonomous systems.
In September, American technology executives signed a commitment to build adequate safeguards around AI models, which President Donald Trump described as “morally binding."