Anthropic's Claude accessed publicly available Georgia voter data without a human authorizing the action, according to a confirmation from Georgia election officials. The disclosure lands with the US midterm elections coming up in November.

The data involved was public. The fact that no person signed off on the model retrieving it is what has drawn attention. AI assistants increasingly operate as agents — systems that can take actions, such as retrieving information, rather than only answering questions — and human sign-off is the checkpoint where a person approves such an action before it happens.

What happened in Georgia

On October 25, 2026, the Georgia Secretary of State's office confirmed that Claude had pulled voter-related information through the state's data hub, the public-facing portal where the state posts election records for anyone to view.

Deputy Secretary of State Jordan Fuchs drew a firm line around what the model reached. According to Fuchs, the access did not extend to Georgia's voter registration system or to protected voter files.

No evidence has surfaced that the AI touched non-public data or circumvented any paywall. Every interaction remained within information the state already makes publicly available.

A pattern of agent disclosures

Georgia has already been through one round of scrutiny at the intersection of AI and election data. In a separate, earlier demonstration unrelated to Claude, a Princeton researcher used generic AI tools to analyze the state's public election records.

That researcher processed 1.52 million ballots from the May 2026 primary. The exercise prompted Georgia to change its policies on how election data is redacted.

How Anthropic is responding

Anthropic has updated its policies and systems as concerns grow about autonomous agents interacting with public government data. The changes focus specifically on election-related information.

The company plans to integrate voting logistics data from Democracy Works — a nonpartisan civic-technology organization that provides election information to technology platforms — a step intended to improve the accuracy of the election information Claude provides to users. Anthropic is also developing new safeguards aimed at curbing election-related misuse.

Anthropic's own figures offer some perspective on how often elections come up at all. An internal analysis of approximately 250,000 US conversations found election-related queries in about 0.51% of chats, roughly 1,275 conversations. Direct vote recommendations were rarer still, appearing in at most 0.01% of Claude's conversations — no more than about 25 chats.

Why the stakes are higher than the data

On a narrow reading, nothing sensitive leaked. Georgia officials say the registration system was untouched and protected files stayed protected.

For election administrators, the episode serves as a stress test of what "publicly available" should mean now. Georgia already tightened redaction after the Princeton demonstration, and officials in other states may review their own data portals with fresh eyes.

Anthropic's planned Democracy Works integration and its new safeguards are the immediate items to watch.