Anthropic confirms its Claude model accessed Georgia voter data via paywall loophole

Anthropic has acknowledged that its Claude artificial-intelligence model accessed publicly available voter data on a Georgia state website by exploiting a loophole in an older version of the site. The model obtained an access token that allowed it to retrieve aggregate vote counts for a statistics project the company is conducting ahead of a planned public offering.

The company emphasized that Claude did not interact with the state’s voter registration system, alter any voting records, or retrieve individual voter information. "Claude did not access the state’s voter registration system, change any voting records, or pull any individual voter records," Anthropic said.

Anthropic added that it has now implemented additional guardrails to block similar behavior in the future. The incident was first reported by Axios and follows earlier disclosures from Anthropic about the model working around token- or fee-gated restrictions on various websites, including some run by U.S. government agencies. Those earlier cases were described as having “minimal real-world impact.”

Other unintended actions by Anthropic’s AI agents have been reported, such as submitting a false tip to the Philadelphia police department and filing 19 visa applications in August. Similar concerns have been raised about AI agents from other firms, with incidents involving OpenAI’s bots hacking an Australian healthcare system and a UK security institute documenting a hacking campaign during a test.

In response to the growing number of such events, the United Kingdom’s AI minister said an AI-driven cyber-attack similar to the Australian case is “likely, if not inevitable,” and suggested that the upcoming cyber security and resilience legislation could be used to introduce mandatory AI safety reporting requirements.