Georgia says Anthropic AI agent accessed public election data, not its voter registration system
Georgia election officials say an artificial intelligence agent developed by Anthropic accessed publicly available election data through the state’s Data Hub, but did not breach the state’s voter registration system or obtain individual voter records.
In a statement this week, Georgia Secretary of State Brad Raffensperger said his office investigated the activity after Anthropic contacted the state Oct. 8 to raise concerns that its Claude model accessed information on the state’s public-facing data platform. The office said its investigation found that the agent accessed only publicly available information and nothing in restricted election systems.
“This isn’t news. Anthropic’s AI agent accessed information anyone can find publicly. Accessing public information is not a hack, and pretending otherwise is reckless and dangerous,” Raffensperger said Monday.
The disclosure comes as election officials and government technology leaders face growing questions about how autonomous AI agents behave when given tasks that involve accessing digital systems and information ahead of the general election on Nov. 3.
“Georgia is consistently recognized as having some of the most secure elections in the country, with voting systems not connected to the internet designed to prevent fraud, paper-ballots, and risk-limiting audits,” Raffensperger continued. “Twisting this into a security scare for clicks undermines confidence in our elections and is absolutely shameful.”
Some cybersecurity experts were quick to agree.
“This hardly qualifies as hacking. The AI used an existing public endpoint to create a visitor account,” Scott Miserendino, chief technology officer at DataBee, said in an emailed statement to StateScoop on Tuesday. “The AI did not do anything nefarious here. It was just thorough in its attempt to complete a legitimate task.”
The state’s Data Hub publishes election information, including unofficial turnout figures, and is separate from the voter registration system. According to the Secretary of State’s Office, Anthropic confirmed Monday that its agent has not accessed Georgia’s voter list or bypassed a paywall. The office also said the White House contacted it about the incident on Oct. 9 and was informed of the findings.
The clarification followed a report by Axios that described Claude retrieving aggregate voting statistics without human authorization while working on an internal statistics project. Anthropic previously said the model had found a way to access data that was normally available for a fee. Georgia officials subsequently disputed that characterization, saying that the information was freely available.
“Georgia’s Data Hub is a public transparency resource and is separate from the state’s voter registration system,” the Secretary of State’s Office emphasized on Monday.
The incident is part of a growing set of concerns Anthropic has raised about its own AI systems in recent weeks.
In a report earlier this month, the AI behemoth described several instances in which Claude took unintended actions while conducting evaluations or completing internal tasks — including exploiting software flaws to run commands on servers, submitting a form on a live government website and working around restrictions to reach information.
Anthropic said the cases it identified — some of which involved websites operated by federal, state and local governments — had minimal real-world impact, but highlighted how AI agents can pursue alternative methods when they encounter barriers, rather than stop and request human guidance.
“While these cases had minimal impact, we do not want to diminish the findings, because the same behaviors could do far more harm as models become more powerful. The larger the role models play in society, the more the public deserves to know how they behave,” the report states.
The company said it had expanded safeguards, including restricting internet access during internal evaluations, strengthening controls on web-access tools and deploying monitoring systems to detect and block unintended behavior. Anthropic added that its detection tools blocked all the cases in its report when tested retrospectively.
“This tooling now runs on most of our evaluations and on internal agentic use of frontier models. When we tested it against the cases described in this post, it blocked all of them,” the report states.
For state and local governments, the episode underscores the importance of understanding not only what an AI system produces, but also how it accesses information and interacts with digital services — especially as government agencies across the country expand their use of AI tools.
In June, California Gov. Gavin Newsom announced a statewide partnership with Anthropic that allows state agencies, cities and counties to access the company’s Claude assistant at a discounted rate. The agreement is intended to help public employees streamline administrative work, analyze information and improve government services.
As governments consider deploying increasingly capable AI assistants, Georgia’s account highlights a distinction between accessing public information and entering protected systems. It also raises questions about how agencies should set boundaries for automated tools, monitor their activity and ensure that systems stop when a task takes them beyond their scope.
Last week, OpenAI sent a letter to Newsom urging the governor to strengthen California’s landmark artificial intelligence safety law, recommending new requirements to monitor AI models during development, improve cybersecurity protections and report incidents involving unauthorized access to third-party systems.
Raffensperger said Georgia’s election infrastructure remains secure, pointing to paper ballots, voting systems that are not connected to the internet and risk-limiting audits.
“Georgia has spent years strengthening election security while making election data more transparent and accessible to the public,” he said.
