Anthropic's Claude accessed Georgia's paid election data without authorization, company says

Anthropic's AI model Claude accessed Georgia's paid vote-tally data without human authorization, the company has disclosed.

Claude did not access the state's voter registration system and did not alter or retrieve any voting records, according to the company. However, the fact that the model bypassed a paywall and accessed voter-related data without human approval has heightened concerns about AI security.

Axios cited an Anthropic spokesperson in reporting the news Sunday, according to Yonhap.

According to the report, Claude accessed vote-tally data that Georgia makes available to the public for a fee while carrying out an internal statistics project. The spokesperson added that the model did not access the state's voter registration system and did not alter or retrieve any voting records.

According to Georgia state officials and Anthropic, the model bypassed a paid-access restriction in the process of retrieving the information.

"The model was being used to gather public data for an internal statistics project," the Anthropic spokesperson said. "Without being instructed to do so, the model requested a token to access a dashboard containing information that is normally available for a fee. We have since put measures in place to prevent this type of behavior."

A source familiar with the matter told Axios that Anthropic had informed the Georgia secretary of state's office and the White House several weeks earlier, and that Trump administration officials had urged the Georgia secretary of state's office to disclose the matter sooner — a request that was not accepted.

Axios noted that Anthropic had previously drawn criticism from the administration for failing to disclose AI-related incidents promptly, and said Georgia now faces similar scrutiny.

Anthropic's AI models have previously been found to act outside human control.

On Oct. 9 (local time), Reuters and AFP reported that an Anthropic AI model had generated a false tip related to a murder case and submitted it to police — a fact that only came to light later.

The AI posed as a person with information about a case, but the tip was classified as spam and never forwarded to the actual investigative unit.

The false tip was submitted on July 18, and Anthropic discovered it around Sept. 28 — roughly two months later — before halting the model's automated testing process, according to reports.

On Oct. 7, Anthropic notified the Philadelphia Police Department of the incident and said it planned to release a report detailing the unintended behavior of its model.

However, no unauthorized access to police systems or data breach was found.

Separately, in July, OpenAI's "rogue agents" caused a stir after breaking out of a sandboxed environment and hacking Hugging Face, an external organization, without authorization. Last month, an OpenAI AI agent was confirmed to have hacked an Australian health data portal.