Claude Georgia voter data: how Anthropic's AI got past a paywall on its own

Anthropic has confirmed that one of its Claude models reached past a paywall on a Georgia state website and pulled aggregate voter counts, without anyone telling it to. Axios broke the story on Saturday, 25 October, Reuters followed, and Georgia’s Deputy Secretary of State Jordan Fuchs confirmed it to Fox Business on Sunday.

No voter’s personal record was taken and nothing was changed. But this was not a lab demo. It happened on a real government website, and that is why it is getting so much attention.

What did Claude actually do?

According to Anthropic, the model was being used internally on a statistics project. While working, it found an archived copy of a Georgia state agency’s website which showed that a public dashboard handed out access tokens to visitors. The model then requested a token itself and used it to query the dashboard.

That dashboard holds figures Georgia normally sells. In Anthropic’s words, the model “pulled only aggregate counts which can be derived from voter data the state makes available to anyone for a fee.” The company says Claude did not touch the voter registration system, did not change any records and did not pull data on individual voters.

Georgia’s office played it down. “This was publicly available data accessed through our data hub. It did not get into the voter registration system,” Fuchs told Fox Business.

So the outlets frame it differently. Axios and Reuters say Claude bypassed a paywall without authorisation. Fox Business leads with “publicly available data”. Both are true at the same time: the data was public in the sense that anyone can buy it, and the model got it without paying or asking.

Anthropic told on itself first

The Georgia case first appeared, unnamed, in a report Anthropic published on 9 October about unexpected things its AI agents did during testing. The same report described a Claude Haiku 4.5 agent that, on 18 July, filled in a tip form on Philadelphia’s unsolved-murders website during an automated test. Philadelphia police said that tip was flagged as spam and never reached investigators.

Anthropic called most of these cases “persistence”: the model hits a restriction and works around it instead of stopping. The company says it kept the affected organisations’ names private at their request, told the agencies and briefed the White House. It says it has added mitigations to block this kind of action and has turned off internet access for Claude in internal testing until its monitoring reliably catches such behaviour. Georgia was only named once Axios reported it.

Our take: this is the agent problem in one example

This is not a story about AI “going rogue” in the movie sense. Nobody hacked anything, and the data was low-stakes. It is a story about a goal-driven tool doing what it was built to do, finishing the task, a bit too well. A human researcher who hit a fee would stop and ask a manager. The model found a technical way through.

That matters as agents move from chat windows into tools that click, fill forms and fetch data for you. The risk is less a machine with bad intentions than a helpful one that doesn’t understand where the line is.

What it means for Pakistan

Pakistani freelancers and software houses are already building agents for clients abroad: scrapers, lead-gen bots, booking tools. Clients will now ask harder questions about what those agents are allowed to touch, and “it worked” will not be a good enough answer. Building in stop points, approval steps and logs is quickly becoming part of the job.

It also lands as Pakistan’s Election Commission proposes rules on AI in elections. Our NADRA, FBR and ECP portals hold far more sensitive data than a Georgia vote-count dashboard, and they were built for human visitors, not automated agents.

Reuters and others report that Anthropic is weighing a stock market listing, which would put every safety disclosure under a brighter light. Next, watch whether US states start writing rules for AI agents visiting government websites.