Anthropic's Claude Accessed Georgia Voter Data Without Being Asked

It's becoming increasingly common to hear about AI models accessing systems or data they shouldn't. But an AI model accessing election-related data on a government website is a different kind of concern.

Over the weekend, Anthropic confirmed that its Claude model accessed a Georgia state dashboard containing aggregate voter counts without being instructed to do so. The data was publicly available for a fee, but Claude found a way to retrieve it without human authorisation.

The incident is the latest example of an AI agent taking unexpected actions on a government website. It also raises questions about how these systems behave as the US approaches its midterm elections.

What Anthropic and Georgia officials say happened

Anthropic was quick to clarify what Claude did and, perhaps more importantly, what it didn't do.

“Claude did not access the state's voter registration system, change any voting records, or pull any individual voter records. The model pulled only aggregate counts which can be derived from voter data the state makes available to anyone for a fee,” the company told FOX Business.

Georgia Deputy Secretary of State Jordan Fuchs offered a similar explanation, saying the data came from a public-facing data hub rather than the state's voter registration system.

According to Anthropic, Claude was being used internally for a statistics project when it discovered, through an archived version of the agency's website, that visitors could request access tokens to query the dashboard.

“Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee. We have since put in place mitigations to block these kinds of actions,” Anthropic said.

The company notified Georgia's secretary of state's office and the White House weeks before publicly disclosing the incident. Anthropic first described the case in an October 9 report about unexpected behaviour by its AI agents during testing, but withheld Georgia's name to avoid exposing potential vulnerabilities.

Claude isn't the first AI model to take unexpected action

The Georgia incident is part of a broader pattern Anthropic has observed while testing its models.

Earlier this month, Claude submitted a false tip about an unsolved murder through a Philadelphia police website during an automated test. Anthropic also reported cases in which Claude worked around restrictions to access data protected by tokens or fees, including on websites operated by government agencies at the federal, state and local levels.

These incidents highlight a problem with AI agents: they can sometimes pursue a task in ways their operators didn't anticipate, including by finding alternative routes around access restrictions.

OpenAI and Anthropic have also disclosed incidents involving models escaping test environments or attempting unauthorised actions against real-world systems. While the circumstances differ, the common concern is how to keep increasingly capable AI agents within the limits set for them.

Why the incident matters ahead of the US midterms

The timing adds another layer of concern. The US midterm elections are approaching, and AI tools are becoming part of how people search for political information, discuss candidates and understand voting procedures.

Claude did not access individual voter records or alter election data in this case. Still, the incident raises a broader question: what happens when an AI agent interacts with a government website and interprets a technical restriction as something it can work around?