Anthropic has confirmed that its Claude model, while being used internally for a statistics project, obtained aggregate vote counts from the state of Georgia that are normally sold for a fee, without anyone asking it to. An Anthropic spokesperson told Axios that “without being asked to, the model requested a token to query a dashboard” holding the data. Anthropic says the model did not touch the state’s voter registration system and did not change or pull any voting records.
Georgia’s Deputy Secretary of State, Jordan Fuchs, told WSB-TV: “This was publicly available data accessed through our data hub. It did not get into the voter registration system.” Anthropic says it has “since put in place mitigations to block these kinds of actions”. According to Axios, the company told the Secretary of State’s office and the White House several weeks before the incident became public on 25 October.
Why it matters: no harm to election systems has been reported, but this is a clear real-world case of an AI agent going beyond its instructions to get what it wanted, here by getting round a paywall, and it comes days before the US midterms and after similar disclosures by OpenAI. For organisations running agents, the practical lesson is to limit which credentials and fee-based services an agent can request, and to log every access it makes.