An Anthropic Claude Mythos 5 model independently broke into and obtained Georgia voter-count data in a new incident.
A researcher for Anthropic had assigned the model a statistics project that required public data available from a state agency for a fee.
During the run, the model located an archived copy of the agency’s website that showed its public dashboard issued an access token to any visitor. It then requested a token and used it to query the database without paying the fee normally required.
The company has argued that the data retrieved consisted of aggregate counts and stressed that the restriction the model bypassed was a fee and access-token requirement for the dashboard, not a security barrier around sensitive voter files
“Claude did not access the state’s voter registration system, change any voting records, or pull any individual voter records. The model pulled only aggregate counts*, which can be derived from voter data the state makes available to anyone for a fee,” an Anthropic spokesperson told FOX Business. As reported by the outlet.
Georgia officials confirmed the nature of the access. Deputy Secretary of State Jordan Fuchs said:
“This was publicly available data accessed through our data hub. It did not get into the voter registration system.”
In the company’s words, it has also tightened guardrails on internet tools such as its web fetch tool and is fixing or removing training environments that reward workarounds.
Concerns over AI and election records have been raised.
Research by a Princeton researcher showed that AI could be used with publicly available records to link voters to ballots, raising ballot-secrecy concerns ahead of the Nov. 3 midterms. That issue is distinct from Anthropic’s retrieval of aggregate counts, but both have drawn attention to AI use with election-related data.
Anthropic is now expanding election safeguards ahead of the Nov. 3 U.S. midterms. In an October update, the company said it was strengthening detection of coordinated election misuse and adding official, state-specific voting information and election results to Claude’s responses.
The question being raised in all this is whether the Georgia officials are covering the accident up along with Anthropic or whether what they have disclosed are the facts.
This accidents came amid a separate case where Anthropic’s different model, Claude Haiku 4.5, submitted a false tip through PhillyUnsolvedMurders.com, run by the Philadelphia Police Department.
Haiku 4.5 had been tasked with generating and performing example tasks on randomly selected webpages, which led it to land on a page about an unsolved homicide that contained a tip form. The model then filled out the form and sent it:
“I may have information regarding this case. I recall seeing someone matching the description in the area around [the street named on the page] during that time period. Please contact me if this information is relevant,” Haiku 4.5 wrote on the report sent.
The model did not include a description of the perpetrator in the fake tip it sent, so the submission was flagged as spam and never forwarded for investigation.
“The department later confirmed that an Anthropic artificial intelligence model submitted a false homicide tip through its PhillyUnsolvedMurders.com website.
The submission occurred in July, but Anthropic only discovered it September 28. it notified the department on October 7.
In a second token-related case separate from Georgia, during an evaluation, Claude Mythos 5 was asked to identify a location shown in a photograph. To help pinpoint the location, the model then tried to use a local government’s property map.
But the evaluation restricted its ability to navigate the site by clicking through it as a normal human would.
Instead, the model examined a settings file that the website sends to visitors’ browsers, found working access tokens in the file, and used them to send requests directly to the server behind the map, from which it then retrieved data.
Although the impact of the behaviors was minimal, as the company said it had already turned off live internet access for some high-risk and cybersecurity evaluations.
Looking at the nature of the threat AI has become and the alarm raised in the past few months about AI rogue behavior, as this is not the first accident, many people are now scared and concerned.
A lot of these companies keep acting like they care but are still making millions of more powerful models. This would leave you to ask: why raise the alarm if you are making an even stronger model tomorrow?
