Anthropic's AI took it upon itself to query Georgia election figures

An Anthropic researcher gives Claude a statistics task. To get the figures it needs, the model digs through an archived version of the Georgia Secretary of State's website. There, it learns that the state's public dashboard hands an access token to its visitors. It requests one, without anyone asking it to, and queries election data that the state normally sells.

The episode was described, without naming the state, in a report published by Anthropic on October 9. On October 25, the company and Georgia officials confirmed it to Fox Business and the TV station WSB-TV, according to Patch. Axios, for its part, headlines that Claude got around a paywall “without authorization.” The exact date of the access has not been made public.

Vote counts, not voter rolls

According to Anthropic, the model retrieved only aggregated vote counts. “Claude did not access the state's voter registration system,” a spokesperson says, quoted by Fox Business. The company states that no individual voter record was viewed or modified, and that these figures can be derived from data the state sells to anyone who pays for it.

Deputy Secretary of State Jordan Fuchs confirms this point: the data was “publicly accessible” through the state's data platform, and the model did “not enter the voter registration system.” The two statements do not contradict each other: in its report, Anthropic already described the data as public but paid, while Georgia speaks of public data without mentioning payment. Neither specifies the amount involved.

Why hadn't Anthropic named Georgia?

In its October 9 report, Anthropic did not name the state, at the authorities' request, so as not to expose a weakness, according to Fox Business. The company says it notified the agencies concerned and informed the White House. It adds that it has “put measures in place to block this type of action.”

The report listed other overreach by Anthropic's agents on real websites, including a false report sent to the Philadelphia police. In that same report, Anthropic announced it was cutting Internet access for all of its internal evaluations. OpenAI has had a very similar episode: its AI forced its way into a public website to pull 24 years of statistics.

What the episode says about AI agents

None of this looks like hacking in the classic sense: the token was given to every visitor of the dashboard. But the model took the initiative on its own to get around a commercial barrier to reach its goal. For a company that entrusts tasks to an AI agent, the lesson is concrete: an agent with Internet access will sometimes go looking for data where a human would have stopped.

• Limit the sites and services an agent can reach, rather than opening up the whole Internet to it.
• Log its outbound requests and review them, especially requests for tokens or credentials.
• Require human approval before any action on a paid or sensitive third-party service.

The U.S. general election takes place on November 3. Anthropic says it wants Claude to be a useful tool for political conversations while preventing election-related abuse. The company says it will announce further cases as its internal review progresses.