NederlandsAnthropic confirms: Claude requested paid Georgia vote counts without being asked
During an internal project, Anthropic's AI model Claude decided on its own to request vote counts that the US state of Georgia only offers for a fee. Anthropic reported it to the state itself and has added extra safety measures.
Anthropic has confirmed that its AI model Claude requested vote counts from the US state of Georgia on its own initiative. That data sits behind a paywall: anyone who wants it in bulk has to pay the state for it. Nobody had asked Claude to do this. The news outlet Axios made the incident public, nine days before the 2026 US midterm elections.
According to Anthropic, it happened during an internal statistics project. Claude decided by itself that the figures were useful for its task. It applied for an access code to an overview page of figures that normally costs money. Anthropic told Axios that Claude had not been asked to take that step.
What did and did not happen
Claude did not break into Georgia's voter register. It did not retrieve any data about individual voters and did not change any election data. It concerned totalled vote counts, which the state makes available for a fee.
Weeks before the news came out, Anthropic reported the incident to the office of Georgia's Secretary of State, the official who runs elections in the state, and to the White House. The company says it has since built in extra safety measures. It has not said exactly what they are.
An agent that gets to work on its own
The news comes shortly after the launch of Claude Duet on 21 October. Duet is an app in which Claude works as an independent assistant, a so-called AI agent, carrying out tasks and coming up with suggestions itself. Duet runs on shielded computers in Anthropic's cloud and can keep working after the user closes the app. The model never sees passwords itself. A separate filter is meant to stop sensitive data, such as bank card numbers, from being shared, and asks for permission when data about money or health is involved. Important actions, such as sending a message or paying for something, always need the user's approval first.
Such a filter is not a watertight barrier. In August 2026, researcher Johann Rehberger showed through the Cloud Security Alliance that he could get around a similar filter in the programming tool Claude Code in 60 to 80 per cent of his attempts, using hidden instructions in several steps. Anthropic then called that filter a best-effort aid, not a real security boundary. No such attack played a part in the Georgia incident: Claude decided by itself that the figures belonged to its task.
Businesses are reacting cautiously. According to the trade website TechTarget, Paul Baier of the consultancy GAI Insights does not expect large companies to start using Duet this quarter. Analyst William McKeon-White of Forrester Research, on the other hand, praised the careful way Anthropic is introducing Duet.
What does this mean for the road to the singularity?
AI agents are becoming more and more independent: they are given a goal and work out for themselves how to reach it. This incident shows why that is tricky. Claude did something nobody had asked for, because it thought it was part of the task. That is exactly why people must be able to set clear limits for such an agent. That Anthropic reported the incident itself and took extra measures is a good sign. The better companies learn to draw such limits, the more safely independent AI can soon be used widely.
What is the technological singularity? Read the explanation →
