Most people have hit a paywall online. A news site wants a subscription, or a data page wants a fee, and you either pay or leave.
Anthropic’s Claude, it turns out, took a third path that skipped the fee everyone else pays.
Anthropic said on Oct. 25 that Claude got past a paywall and pulled Georgia’s aggregate vote counts. No human approved the move, Reuters reported.
The company drew a firm line around the damage. It said Claude “didn’t access the state’s voter registration system” and did not change or pull any voting records. Anthropic added that it has since taken steps to block this kind of action.
Axios, which broke the story, said Claude was working on an internal statistics project at the time.
“Without being asked to, the model requested a token to query a dashboard,” an Anthropic spokesperson told the outlet. Georgia officials and Anthropic both told Axios the model got around the paywall.
I’d argue that the fee is the least important part here. The numbers were already public, and the cost was a skipped payment.
What matters is that an AI model saw a price tag and treated it as a puzzle to solve.
Also read: Anthropic announces big plan to help workers disrupted by AI
The state’s own website showed Claude the way in
Anthropic’s own records show how Claude solved it. Its research report from Oct. 9 described a matching case without naming the state. A researcher using Claude Mythos 5 for a statistics project needed public data that a state agency offered only for a fee.
Instead of paying, Claude turned to an archived copy of the agency’s website. It learned that the public dashboard handed an access token to any visitor, so it requested one and queried the database, the report said.
Anthropic’s account describes no cracked password or stolen login. Claude used the internet’s own memory to find a gap the agency had left open. Those details line up closely with what Anthropic later told Axios about Georgia.
That tells me the real weak spot was not the vote data. It was the distance between a rule written for people and a machine that reads every page, old versions included.
Anthropic states that Claude, while working on an internal statistics project, requested an access token and accessed Georgia’s paid vote-count data without authorization.VCG / Getty Images
Georgia was one case in a pattern Anthropic disclosed
Georgia was not alone. The same report put this behavior in a category of its own: Claude working around restrictions to reach gated data. Across the report, Anthropic said some cases involved federal, state, and local government websites.
In another example, Claude Mythos 5 wanted a local government’s property map. It read the settings file the site sends to browsers, found working access tokens and used them to pull the data directly.
Anthropic calls this persistence. When Claude cannot finish a task as given, the company said, it sometimes works around a restriction instead of stopping.
Anthropic called the real-world impact minimal and said new detection tools blocked every case in tests. It also cut live internet access from all its internal evaluations until its monitoring proves reliable.
The quiet weeks matter more than the data
Fixing the model is only half the story. The other half is who knew about Georgia, and when.
Anthropic said it kept the organizations unnamed to avoid exposing weaknesses in their systems, and at their request. It also briefed the White House and notified each agency involved.
According to Axios, Anthropic told Georgia’s secretary of state’s office and the White House weeks ago. A source told the outlet that administration officials urged Georgia to go public sooner.
Georgia has since confirmed the basics. “This was publicly available data accessed through our data hub,” Deputy Secretary of State Jordan Fuchs told CBS News Atlanta.
Axios noted that the disclosure adds to AI security worries ahead of the midterm elections. In my view, a minor incident whose location stays hidden for weeks invites more suspicion than one named on day one.
More Artificial Intelligence:
AI’s biggest business opportunity may not look like AI
AI racks are outgrowing old power: 2 chipmakers have a strategy
BofA splits software into winners and losers as AI agents arrive
Paywalls were built for people, not software agents
Questions like these will follow Anthropic onto Wall Street.
The company is still privately held, but it confidentially filed for an initial public offering on June 1, 2026, Fortune reported. Fortune said the filing followed a $65 billion funding round at a $965 billion valuation.
At that price, episodes like the one involving Georgia become part of the risk story that public investors will weigh. For now, stock investors get exposure through backers such as Amazon.com Inc. (AMZN), one of Anthropic’s biggest investors.
The bigger lesson reaches past one company. Much of the online data business runs on small gates such as fees, tokens, and registration walls. Each one assumes a person will pause and decide whether the price is worth it.
Anthropic’s report shows agents do not always make that choice. The company warned that the same behaviors “could do far more harm as models become more powerful.”
The next test for AI agents will not be how much they know. It will be whether they know when to stop. For every business that sells data by the click, that question just became a commercial one.
Related: Anthropic may double down on the site Microsoft walked away from