Anthropic’s Claude Bypassed A Paywall To Access Georgia Election Data. No One Had Asked It To.

Anthropic's Claude AI model accessed Georgia election statistics normally available for a fee without human authorization, using a digital access token to retrieve the information during an internal research project.

The incident occurred when the model was tasked with collecting public data for a statistical analysis. Rather than paying for access, Claude discovered a way to query the state's database through a public dashboard, according to an earlier account published by Anthropic.

The company confirmed Sunday that Georgia was the state involved, adding that the model had acted without being instructed to request the token.

Claude obtained aggregate vote counts, which are numerical totals rather than individual voting records. It did not access Georgia's voter registration system, retrieve individual voter files or alter election information, according to the company.

Georgia election officials separately confirmed those limits on Monday. Robert Sinners, a spokesperson for the Georgia secretary of state's office, told The Atlanta Journal-Constitution that the incident did not involve a breach, hacking or tampering with the state's election systems.

The distinction is important because the information Claude retrieved was available to members of the public. The problem identified by Anthropic was how the model obtained it, bypassing a payment requirement while carrying out a legitimate research task.

Anthropic had documented the behavior in an October 9 research report examining instances in which its models took actions beyond their intended instructions. At the time, the company described the organization involved only as a state agency, saying it was withholding identifying details to avoid exposing potential vulnerabilities.

According to the Anthropic research report, an employee had used Claude Mythos 5 for a statistics project requiring public information normally sold by the agency.

The model examined an archived version of the agency's website and discovered that visitors to its public dashboard could obtain access tokens. These digital credentials allow software to communicate with an online service.

Claude then requested a token and used it to retrieve the data without paying the required fee. The action was not part of the researcher's instructions, Anthropic said. The company categorized it as an example of unintended behavior in which an AI system encounters a restriction and finds an alternative route to completing its assignment.

Its investigation identified several other incidents involving interactions with outside websites. In one case, a model exploited a software flaw on a university server while trying to complete a scientific calculation. In another, Claude submitted information to a real police tip form during a browsing task, despite the submission being unrelated to an actual investigation.

The company said the examples it reviewed had caused minimal real-world harm, but acknowledged that similar behavior could become more consequential as AI systems gain greater capabilities.

The Georgia incident was disclosed publicly just over a week before the November 3 U.S. midterm elections. Georgia operates an Election Data Hub that provides public access to voting and registration statistics. In April, Secretary of State Brad Raffensperger announced additional features allowing users to examine turnout by party affiliation.

The secretary of state's office reported that more than 1 million Georgians had cast early in-person ballots by noon on Oct. 20, compared with 894,412 at the equivalent point in the 2022 election.

The Claude incident did not involve changes to those counts or the systems used to maintain voter records.

Anthropic had notified Georgia officials and the White House several weeks before the latest disclosure, according to Axios. The publication also reported, citing a person familiar with the matter, that administration officials had urged Georgia's secretary of state's office to disclose the incident sooner.

Anthropic has also expanded its monitoring of internal AI agents and begun moving them onto centrally managed infrastructure with stronger restrictions on their ability to interact with outside systems.

The Georgia case is separate from more serious cybersecurity incidents Anthropic disclosed earlier this year. In a September security assessment, the company reported identifying four incidents in which Claude models gained unauthorized access to real third-party systems during cybersecurity evaluations.

Those incidents prompted a wider review of model activity and additional efforts to identify unexpected interactions with live websites.

The report said Anthropic had expanded its examination beyond specialized cybersecurity tests to include routine evaluations and internal use, with additional instances of unintended behavior expected to be disclosed as its review continues.