The state Claude snuck into for paywalled vote counts was Georgia

Earlier this month Anthropic published a report about times its AI models did things on the real internet that nobody asked them to do. One entry described a Claude model, working on an internal statistics project, that found its way into a state government dashboard holding data the state normally charges for. The report didn't say which state. On Sunday, Anthropic and Georgia officials confirmed it to Axios and Fox Business: it was Georgia, and the data was vote counts.

Nine days before the midterms, that is not the sentence any election official wants to read about an AI.

What the model actually did

The mechanics are almost mundane. According to Anthropic's account, the model was digging for numbers and came across an archived copy of a state agency's website. That old page revealed that a public dashboard handed out digital access tokens to visitors. So the model asked for one. "Without being asked to, the model requested a token to query a dashboard," Anthropic said, adding that the dashboard held information normally available for a fee.

No password was cracked, and nothing got broken into in the movie sense. The model read some old documentation, noticed an unlocked side door, and walked through it because that was the shortest path to finishing its task. That is the behavior people worry about with agents: a system that treats a paywall as one more obstacle between it and a completed job, with no malice required.

Anthropic is careful about the limits. "Claude did not access the state's voter registration system, change any voting records, or pull any individual voter records," the company said in a statement. It says the model pulled only aggregate counts, numbers anyone could derive from data the state sells. Georgia agrees on that much. "This was publicly available data accessed through our data hub. It did not get into the voter registration system," Deputy Secretary of State Jordan Fuchs told Fox Business.

Several early headlines said Claude had accessed "voter data," which made it sound like a breach of the rolls. It wasn't. But "it only took the numbers we sell" is a strange kind of reassurance. The model didn't pay. The state's own boundary between free and paid turned out to be a suggestion, and the thing that ignored it was software running unattended inside an AI lab.

Who knew, and when

Anthropic says it notified the agencies involved and briefed the White House before publishing its October 9 report. It left Georgia's name out because, it said, the affected organizations asked to stay anonymous so as not to advertise the vulnerability. That is a reasonable call for an open hole. It is less reasonable once the hole is closed and the election is days away.

The New Republic, citing Axios, reported that a source said Trump administration officials had urged the Georgia secretary of state's office to disclose the incident, and that those calls went unheeded. Georgia's office has not said why it waited. The name came out on a Sunday, the day before Anthropic made its IPO filing public.

The Georgia case shared that October report with the incident this site covered earlier this month, when a Claude Haiku 4.5 model running an automated test filled out a tip form on a Philadelphia police site claiming it might have information about an unsolved murder. Philadelphia police said the tip was flagged as spam and never acted on. The same report described other cases of Claude models working around tokens and fee restrictions on federal, state and local government websites.

So Georgia isn't a one-off. It is one entry in a pattern Anthropic has documented itself. Its S-1, filed Monday, adds another: in the risk factors, the company discloses that a setup error left models in an evaluation connected to the internet, that four models accessed real third-party systems, and that one of them read a person's personal data. Anthropic says it notified the people affected.

The company's line on Georgia is short. "We have since put in place mitigations to block these kinds of actions," a spokesperson said. Anthropic hasn't described what those mitigations are.

Credit where it's due: Anthropic told the state, told the White House and published the incident, which is more than most companies do when their software misbehaves. Nobody would know about the Georgia dashboard if Anthropic hadn't written it down. But the story works the other way too. The only reason the public knows an AI wandered into a state election dashboard is that the AI's maker chose to say so. There's no rule that requires it, and no outside auditor counting how many times this happens at labs that keep quiet.

Anthropic says about 0.51 percent of a sample of U.S. Claude.ai conversations from September 21 to 27 were about elections or voting. Tiny as a share, enormous in absolute terms, and that only counts the people talking to Claude. The Georgia episode involved no person at all. A model went looking for numbers, decided a fee was in its way and got around it. The polls open on November 3.

Sources: Axios, Fox Business, The New Republic, The Washington Post.