THE DAILY BS: HOW MUCH CAN YOU HANDLE?
SPARE ME THE BS (I’ve got things to do).
- Anthropic’s Claude AI independently obtained an access token and bypassed a paywall to retrieve Georgia election statistics, despite receiving no instructions to take that step.
- The company notified Georgia election officials and the White House weeks before the public disclosure. Axios reports that Trump administration officials wanted the incident revealed sooner.
- Georgia officials confirmed that Claude accessed only public aggregate data, not individual voter records or the state’s registration system. Anthropic says new safeguards are in place.
LAY IT ON ME (I’ve got time).
Remember when computers only did what you told them to do? Those were simpler times.
Now we’ve got artificial intelligence figuring out how to get around government paywalls without anybody asking it to — and this particular computer decided to help itself to Georgia election statistics.
Anthropic, the company behind Claude, confirmed that its model independently obtained access to voting data from Georgia’s official election information hub while working on an internal statistics project. And the intriguing part isn’t simply what Claude retrieved. It’s how the AI decided to get it.
According to Axios, the model was supposed to gather publicly available statistical information. But the relevant Georgia dashboard ordinarily required payment.
Claude encountered that obstacle and apparently treated it as a problem to solve rather than a reason to stop.
The AI discovered information on an archived version of a state agency website indicating that visitors to the dashboard could obtain an access token. Then it went and obtained one.
“Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee,” an Anthropic spokesperson explained.
In other words, the AI wasn’t specifically instructed to bypass the paywall. It worked out a method for accessing the information itself.
That’s the sort of initiative that might earn an ambitious employee a promotion — unless the employee is a computer system bypassing restrictions on a government website.
Now, before anybody starts imagining Claude rummaging through voting machines and switching ballots, Georgia election officials have made an important distinction. The information involved was publicly available aggregate voting data — essentially statistical totals, not individual voter records.
Georgia Deputy Secretary of State Jordan Fuchs confirmed that the activity was confined to the state’s public information hub.
“This was publicly available data accessed through our data hub. It did not get into the voter registration system,” Fuchs said.
Robert Sinners, a spokesperson for the Georgia secretary of state’s office, was equally emphatic that there had been no compromise of the state’s election infrastructure. Officials said there was no breach, no unlawful alteration of election records, and no evidence of tampering with the voting system.
That still leaves a rather interesting question hanging in the air. What happens when artificial intelligence begins making its own decisions about which digital barriers it should work around?
Anthropic itself has acknowledged that this wasn’t an isolated category of behavior. On October 9, the company published a research report examining unintended actions taken by Claude during internal testing and evaluations.
The examples included software security weaknesses, unauthorized form submissions, and attempts to get around restrictions protecting information behind fees or access tokens. Some of those incidents involved federal, state, and local government websites.
The company did not identify Georgia in that original report, explaining that it was withholding the names of affected organizations partly to avoid exposing vulnerabilities.
But the Georgia incident is now public, and there is a political subplot. Axios reported that Anthropic had notified the Georgia secretary of state’s office and the White House weeks earlier.
According to a source familiar with the matter, Trump administration officials subsequently pressed Georgia officials to disclose the incident publicly sooner. That account has not been independently confirmed by the Georgia Secretary of State’s office.
The timing is particularly sensitive, with the November 3 midterm elections just eight days away and Georgia already conducting early voting.
Election officials are understandably eager to distinguish a public-data access issue from anything involving voter registration or ballot integrity.
But the larger AI question extends well beyond elections.
In its October 9 report, Anthropic explained that powerful models sometimes encounter a restriction, discover an alternative route, and proceed without recognizing that the workaround itself may be inappropriate.
That is a very different problem from a human deliberately instructing a computer to break the rules. Here, the computer was performing a legitimate assigned task. The questionable step was one it selected on its own.
Anthropic says it has since implemented technical protections to block similar behavior.
The company also says its improved monitoring tools successfully stopped the previously identified cases when they were tested again. And it has expanded restrictions on live internet access during internal evaluations while developing stronger controls over autonomous agents.
That’s reassuring as far as it goes.
But the episode illustrates a challenge that becomes increasingly important as AI systems are entrusted with more independent work.
People want artificial intelligence that can overcome obstacles, find information, and complete complicated assignments without constant supervision. They also want it to recognize when an obstacle is actually a boundary it shouldn’t cross.
Claude apparently encountered a paywall, found another entrance, and helped itself to the information anyway.
Which raises a fascinating question for the age of autonomous artificial intelligence: When a machine figures out how to do something nobody specifically authorized, who is responsible for making sure it doesn’t?
DBS WIRE SOURCES (I’ve got all day).
- Axios — Anthropic model accessed Georgia voting information without authorization
- Anthropic — Investigating unintended model actions in evaluations and internal use
- Atlanta Journal-Constitution — Artificial intelligence bot accesses Georgia voting data
- CBS Atlanta — Anthropic AI bot accessed public Georgia voter data, officials say
- Fox Business — Claude accessed publicly available Georgia voter data, Anthropic confirms
