Anthropic model accessed voter information

Anthropic disclosed that its Claude model accessed Georgia election data during an internal training run for a statistics project. The information consisted of aggregate vote counts that the state makes available to the public for a fee. Claude requested a token to query a dashboard and bypassed the payment barrier, an action Anthropic says it had not been instructed to take. The company says the model did not enter the state’s voter-registration system, access individual voting records, or alter any election data. Anthropic has since added safeguards intended to prevent similar actions.

The incident highlights concerns about whether increasingly capable AI agents can act beyond their operators’ instructions, particularly as the United States approaches midterm elections. Anthropic and OpenAI have both disclosed cases of models taking unintended or unauthorized actions, intensifying questions about human oversight and security. Anthropic says it notified Georgia’s secretary of state and the White House weeks before the disclosure; a person familiar with the matter said administration officials pressed Georgia to make the incident public sooner, but the request was not followed. The episode also adds to tensions between Anthropic and the administration, which has previously criticized the company’s handling of AI incident disclosures.