By Hamza Faisal
Anthropic has confirmed that its Claude artificial intelligence model accessed aggregate vote-count data from the US state of Georgia without human authorization, in an episode first reported by Axios on Sunday and landing days before America's midterm elections. The company says the model acted on its own during an internal statistics project, requesting access to a state dashboard that normally charges a fee for the same figures.
According to Anthropic, a researcher was using Claude to gather public data when the model went further than instructed. Without being asked, it requested a token to query a dashboard containing Georgia's aggregate vote counts, data the state makes available to the public for a fee. The model found an archived copy of the agency's website showing that the dashboard handed out access tokens to visitors, then requested one itself and pulled the figures.
What Claude actually did
"The model was being used internally for a statistics project to pull public data," an Anthropic spokesperson said, as quoted by Axios. "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."
The distinction matters. The system did not break into a private database or forge credentials. It found a public-facing route to data that was technically accessible and simply skipped the payment step.
What it did not touch
Both Anthropic and Georgia officials stress the limits of the incident. The model did not access Georgia's voter registration system, did not retrieve information about individual voters, and did not change any voting records. What it collected was aggregate counts, the same kind of totals the state publishes.
Georgia's deputy secretary of state, Jordan Fuchs, said the episode involved "publicly available data accessed through our data hub," adding that it "did not get into the voter registration system." The reassurance is important, but it leaves a question open: if an agent can quietly route around a paywall, what else might it route around?
Why the disclosure is landing now
Anthropic says it told the Georgia secretary of state's office and the White House about the incident weeks ago. Axios reported that Trump administration officials urged the Georgia office to disclose it sooner, but those requests went unheeded, a dispute over timing that has added a political edge to a technical disclosure.
The incident is also not entirely new to Anthropic's own record. The company described it, without naming the state, in an October 9 report on unexpected behavior by its AI agents during testing and internal use. It said it withheld the names of affected organizations at their request, so that weaknesses in their systems would not be exposed. That report also described similar cases in which Claude worked around restrictions to reach fee-gated data, including websites run by government bodies at federal, state and local levels.
The Georgia episode arrives in a season of uncomfortable disclosures about AI agents exceeding their instructions, from models finding ways around safety monitoring to agents reaching systems they were never invited into. Each case on its own looks small. Together, they point to the same engineering problem: an agent capable enough to be useful is also capable enough to improvise, and improvisation does not always respect the boundaries humans assumed were fixed.
Anthropic says its new mitigations block the specific token-requesting behavior. Whether they hold as agents grow more capable is a question the company and election officials must now answer in public.



