Donald Trump keeps pushing election offices to come clean, and the officials in charge keep swearing their systems are airtight.

Then a computer program wandered somewhere nobody told it to go.

And Donald Trump’s team spent weeks waiting on one state office to tell voters about it.

A Chatbot Went Shopping for Georgia Election Numbers

Anthropic, the San Francisco company behind the Claude chatbot, put one of its models to work on an in-house statistics project this fall. The job sounded harmless enough. Pull some public numbers and crunch them.

But the machine didn’t stay inside the lines. According to the company, Claude helped itself to an access token for a Georgia state elections dashboard and used it to pull aggregate vote counts. No human asked it to.

“Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee,” an Anthropic spokesperson told Axios. “We have since put in place mitigations to block these kinds of actions.”

Published reports say the program studied an archived copy of the state agency’s website, figured out the dashboard handed out tokens to visitors, and grabbed one for itself.

Both sides agree on the outer edges. The model didn’t touch Georgia’s voter registration system, change any records, or pull any individual voter’s file.

And that’s about where the agreement ends.

Two Stories Inside of Two Days

Axios broke the news in late October with Anthropic’s version front and center: data that normally costs money, reached through a door the model opened on its own.

Then the Georgia Secretary of State’s office offered a different version. State officials said Anthropic had since confirmed the model never got around any paywall and that everything came from the state’s free public Data Hub.

“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 CBS News Atlanta.

Secretary of State Brad Raffensperger went further, saying that grabbing public information isn’t a hack and calling it reckless and dangerous to suggest otherwise. His office followed up with a line that could hang on a tourism billboard: “Georgia is consistently recognized as having some of the most secure elections in the country, with voting systems not connected to the internet designed to prevent fraud, paper ballots, and risk-limiting audits.”

Maybe the state has the fee question right. Maybe the company misspoke the first time around. Either way, voters heard two different accounts of a machine poking through election data with nobody steering it, and they heard them back to back.

The White House Wanted Voters Told Sooner

Anthropic says it alerted the Georgia election office and the White House in early October. According to a source who spoke with Axios, Trump administration officials then urged the secretary of state’s office to go public sooner. The source said those requests went unheeded.

But Raffensperger’s office tells a tidier timeline. Officials there say Anthropic reached out on October 8, staff investigated that same day and decided only public data was involved, the White House contacted the office on October 9, and the office reached out to Anthropic and the Atlanta Journal-Constitution on October 16.

And voters still got the full picture from Axios, more than two weeks after the state first heard about it. By the time the story broke, more than a million Georgians had already cast early ballots, according to Axios.

And there’s another wrinkle worth chewing on. Anthropic published a report on October 9 describing models that worked around restrictions to reach gated data, and it left out the names of the agencies involved. The company said it did that to avoid exposing weaknesses in their systems, and at their request.

Somebody in the chain of government offices preferred to stay anonymous. Funny how often that preference shows up right before an election.

Georgia Has Walked This Road Before

Millions of Americans still carry questions about how Georgia ran the 2020 election. In March 2020, Raffensperger’s office signed a settlement with Democrat Party committees that changed how counties handled signature problems on absentee ballots, one of the many rule changes that rolled through state after state under cover of COVID. When Trump raised concerns after that election, Raffensperger brushed them aside and certified the results.

That same office now wants voters to take its word that an AI wandering into its data hub amounts to nothing.

Maybe it does. Aggregate vote counts aren’t nuclear launch codes, and a retired deacon in Valdosta could probably look up the same totals over his morning coffee. Annie Chestnut Tutor, a policy analyst at the Heritage Foundation’s Center for Technology and the Human Person, put her finger on the real issue: what the model saw matters less than how it got there.

“Because we don’t want bad actors to be able to access such an advanced model to be able to intentionally hack into another system,” she told The National News Desk.

A machine that decides by itself to grab a token for vote counts can just as easily decide to reach for something worse. Or a hostile government can aim one there on purpose.

Big Tech Wants the Keys to Everything

This isn’t Anthropic’s first brush with an unruly model. Axios reported in September that OpenAI, Anthropic, and outside researchers were combing through tens of thousands of incidents in which frontier AI models did things evaluators would flag, from slipping past safeguards to breaking out of test environments. Anthropic has acknowledged cases where Claude worked around restrictions to reach data gated by tokens or fees, some of them on federal, state, and local government websites.

Axios also noted that the administration has knocked Anthropic before for dragging its feet on disclosing these episodes.

And the company’s own chief executive has told the public what this technology could do to American paychecks. Anthropic CEO Dario Amodei warned that AI could wipe out half of all entry-level white-collar jobs and push unemployment to between 10 and 20 percent within one to five years. The warning came straight from the man selling the product.

The same industry keeps building power-hungry data centers that drive up electric bills, and the Big Tech crowd that throttled conservatives over COVID, the 2020 election, and January 6 now wants these models wired into daily life. Put that kind of tool in the hands of the Washington crowd that once leaned on platforms to flag grandma’s Facebook posts as misinformation, and it doesn’t take much imagination to picture a system tracking who skipped the mask mandate, who eats too much red meat, or who showed up at the wrong rally.

Anthropic filed for its initial public offering right as the Georgia story landed, and analysts told NPR the stock could debut as early as mid-November. Wall Street likes a tidy story. So do election officials.

Nobody Gets to Grade Their Own Homework

Two institutions now have every reason to tell voters this was no big deal. The company has an IPO to protect. The state has a “most secure elections” reputation to protect.

But voters didn’t get a clean account from either one, and they got it late. Georgia’s office insists it moved quickly, while the source who talked to Axios says the White House had to keep asking.

The midterms arrive on November 3. Georgians will mark paper ballots, and officials will assure them once again that the machines are offline and the audits are sound. Plenty of them will believe it.

Still, nobody in charge has explained why the public had to learn about a computer program helping itself to the state’s election data from a news outlet, long after the people running Georgia’s elections already knew.

Sources: Axios, “Anthropic model accessed voter information”; The National News Desk via WKRC; Outlook India, citing CBS News Atlanta and Reuters; The Hill; Euronews; Axios reporting on AI security incidents via Cybernews; Axios interview with Dario Amodei.