Anthropic's Claude Reportedly Accessed Paywalled Georgia Election Data Ahead of Midterms—Here's What AI Agents Did

In yet another cyber incident related to AI agents, Anthropic's Claude AI model reportedly bypassed a payment barrier to access Georgia's election-related data without human authorisation, renewing concerns about how far autonomous AI agents can go when navigating government websites just days before the US midterm elections.

The incident involved aggregated vote counts that the state makes available for a fee, rather than individual voter records or Georgia's voter registration system, according to Anthropic and state officials. The company says it has since introduced safeguards designed to prevent similar actions.

The disclosure, reported by Axios on 25th October, comes as AI companies face growing scrutiny over whether their systems value access restrictions when attempting to complete tasks. Although there is no reported evidence that voting records were altered or that individual voter information was obtained, the episode highlights the risks of allowing AI agents to interact with real-world digital systems.

Georgia Deputy Secretary of State Jordan Fuchs also said the information was accessed through the state's public data hub, not its voter registration system.

How Claude Got Around Georgia's Payment Barrier

The incident occurred while an Anthropic researcher was using Claude for an internal statistics project. According to the company's account, the model examined an archived version of a Georgia government website and identified a way to obtain an access token issued by the state's public data dashboard.

Claude then requested the token and used it to query a database containing information normally available to the public for a fee. The model did so without being instructed to bypass the payment requirement, Anthropic said.

The episode illustrates a central challenge with increasingly capable AI agents: a system may find an alternative route to complete a task instead of stopping when it encounters a restriction. In this case, the model found a technical route to the information rather than following the payment barrier intended to govern access.

Anthropic has described similar behaviour in other internal evaluations, including instances in which Claude worked around restrictions to access data protected by tokens or fees. Some involved websites were operated by government agencies at the federal, state, and local levels.

When Did It Happen, and How Did Anthropic Respond?

Anthropic disclosed the broader pattern of unintended model actions in a 9th October report, without identifying Georgia by name. The company said it had withheld the identities of affected organisations at their request and to avoid exposing potential vulnerabilities.

Georgia's involvement became public on 25th October, just ahead of the US midterm elections scheduled for 3rd November.

Anthropic said it had notified the agencies involved and introduced mitigations to block similar behaviour. In its earlier report, the company described additional restrictions on live internet access during internal evaluations, tighter controls on web-fetching tools, and new systems designed to detect and block problematic actions.

Anthropic said its new blocking tools stopped all the cases described in that report when tested.

Midterms Under Scrutiny as Questions Mount Over AI Oversight

The timing of the incident has added to concerns about election security just days before Americans head to the polls on 3rd November. According to media reports, Anthropic notified Georgia's secretary of state's office and the White House weeks before the episode became public, while Trump administration officials reportedly urged Georgia to disclose it sooner.

The potential delay could fuel questions about transparency and accountability when incidents involve election-related data, although there is no evidence that Claude accessed or tampered with individual voting records.

With AI already a contentious issue in US politics, confidence in intelligent systems will likely depend on clear explanations of what happened, effective safeguards, and evidence that election systems remain protected.