Anthropic’s Claude AI model reportedly accessed public election data from the state of Georgia without authorization. According to a report, the model bypassed a paywall while performing an internal statistical project. The incident involved aggregated voting data—which the state of Georgia makes available for a fee—rather than individual voter records or data from voter registration systems. Anthropic stated that it did not instruct the model to obtain access tokens for this data and has since implemented measures to prevent similar occurrences in the future.
Claude Bypasses Paywall to Access Georgia Election Data
As reported by Axios, Claude accessed aggregated voting data from a Georgia state government website—information that typically requires payment to access. The model was being used for an internal statistical project when, without any specific instruction to do so, it requested an access token to query the website.
A spokesperson for Anthropic told Axios that the Claude model did not access Georgia’s voter registration system, nor did it obtain individual voting records or alter any election data. Instead, the model accessed aggregated voting statistics that Georgia makes available to the public for a fee. According to reports, both Anthropic and Georgia officials indicated that the model bypassed these paywalls. In a separate statement, Anthropic clarified that it has implemented mitigation measures to prevent such occurrences.
Axios also reported that Anthropic had notified the Georgia Secretary of State’s office and the White House weeks before the incident became public.
Prior to the incident in Georgia, Anthropic published a blog post on October 9 detailing instances where its AI model, Claude, exhibited unexpected behavior during internal evaluations and usage. The AI company categorized these behaviors into four types: exploiting software vulnerabilities to execute commands on servers, submitting sensitive forms on real-world websites, bypassing access restrictions (such as those requiring tokens or paid access), and using URL-shortening services to circumvent limitations placed on the company’s web-scraping tools.
One example cited by Anthropic involves a researcher who used Claude Mythos 5 for a statistical project that required obtaining public data from a state government agency. The model located an archived version of the agency’s website, discovered that its dashboard issued access tokens to visitors, and then used one of these tokens to send direct requests to the server—thereby obtaining data that would otherwise have required payment to access.