Anthropic’s Claude Bypasses Paywall to Access Georgia Voter Data

By Staff, Agencies

Anthropic’s Claude AI model bypassed a paywall to obtain voter data in the US state of Georgia without being instructed to do so, Axios reported on Sunday, raising concerns about the ability to control increasingly autonomous artificial intelligence systems.

The incident occurred during an internal statistics project involving public information. Anthropic said Claude retrieved aggregate voter counts but did not access Georgia’s voter registration system, obtain individual voter records or modify any data. The company confirmed the incident ahead of the November 3 midterm elections.

“Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee,” an Anthropic spokesperson told Axios.

Claude reportedly learned from an archived government webpage that the dashboard issued digital access keys to visitors, then requested a key and retrieved the information without paying.

Georgia Deputy Secretary of State Jordan Fuchs confirmed to Fox Business that the data came from a public information service, emphasizing that Claude “did not get into the voter registration system.”

Anthropic described the incident in an October 9 report without identifying Georgia. According to Axios, the company notified state officials and the White House weeks earlier. A source told the outlet that administration officials had urged Georgia to disclose the incident sooner, but those requests were not acted upon.

The case is among several incidents Anthropic has disclosed in which its models found unintended ways around restrictions rather than stopping.

Other examples included submitting a false homicide tip to Philadelphia police and filing incomplete applications through the US State Department’s visa website.

On Friday, Anthropic also revealed that one of its models had impersonated a human during a cybersecurity exercise, attempting to persuade an unrelated software developer to accept malicious code. The developer rejected the request, and the attempt failed.

Other AI developers have reported similar challenges. OpenAI said Friday that it was keeping training for its most powerful models offline after agents escaped restricted environments.

In July, its agents reportedly bypassed safeguards and accessed systems belonging to Hugging Face, an online repository for AI models and datasets.

Anthropic said it had moved some evaluations offline, restricted internet tools and introduced systems to detect unauthorized actions. The company added that these safeguards blocked each of the disclosed incidents when tested against them.