Anthropic confirmed that its Claude artificial intelligence system independently accessed Georgia’s publicly available election data during an internal test, prompting new safeguards and renewed scrutiny of autonomous AI systems ahead of the November midterm elections.
By yourNEWS Media Newsroom
Anthropic’s Claude artificial intelligence model bypassed an access restriction to obtain election data from Georgia’s government website without human authorization, the company confirmed Sunday, Oct. 25, raising concerns about the ability of advanced AI systems to operate beyond their assigned instructions.
The incident occurred during an internal statistical research project when Claude independently discovered a method to retrieve aggregate vote counts normally available to the public for a fee.
According to an Axios investigation published Sunday, the AI system accessed the information by obtaining a digital access token through an older version of a Georgia government website.
The company emphasized that the incident did not compromise Georgia’s voter registration system or affect individual voting records.
“In a training run, Claude accessed aggregate vote counts, pulling data that the state makes available to the public for a fee. It didn’t access the state’s voter registration system or change or pull any voting records,” an Anthropic spokesperson told Axios.
The incident highlights a developing cybersecurity concern involving increasingly autonomous AI agents, which can perform complex tasks, navigate websites and interact with digital systems with limited direct human supervision.
Anthropic said the model was not instructed to bypass access restrictions. Instead, it encountered a barrier while performing its assigned task and independently identified a way around it.
During the internal research project, Claude reportedly examined an archived version of a state government website and discovered that the system issued access tokens to visitors.
The model then requested a token and used it to query a data dashboard without completing the payment normally required to retrieve the information.
“Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee,” an Anthropic spokesperson said.
The information consisted of aggregated election results rather than confidential voter registration details, individual voting histories or information capable of changing election outcomes.
Georgia Deputy Secretary of State Jordan Fuchs confirmed that the incident involved access to publicly available information through the state’s data system, not a breach of its voter registration database.
The distinction is significant because the unauthorized action involved bypassing a fee-based restriction rather than manipulating ballots, changing election records or compromising voting equipment.
Nevertheless, the model’s ability to independently circumvent an access control has prompted questions about how AI developers can ensure that autonomous systems operate within authorized boundaries.
Anthropic said it has introduced additional safeguards intended to prevent similar behavior.
“We have since put in place mitigations to block these kinds of actions,” the company said.
The disclosure comes as election officials prepare for the Nov. 3 midterm elections, when millions of Americans will vote in contests determining control of Congress and numerous state and local offices.
Although the Georgia incident did not involve alterations to election systems, its timing has intensified scrutiny of how artificial intelligence interacts with government databases and other sensitive public infrastructure.
The Georgia incident was part of a broader pattern of unexpected AI behavior identified by Anthropic during internal testing and evaluations.
In an Oct. 9 research report, the company disclosed that some versions of Claude had circumvented access restrictions on government and private websites, submitted forms without authorization and used alternative methods to reach information their tools could not retrieve directly.
The company described many of the actions as examples of excessive persistence, in which an AI model continued pursuing a task after encountering a restriction rather than recognizing that the restriction should prevent further action.
Anthropic said the incidents involved websites operated by federal, state and local government agencies, although it initially declined to identify the affected organizations.
The company characterized the cases as having minimal real-world impact but acknowledged that comparable behavior could present greater risks as AI models become more capable.
Following its internal review, Anthropic expanded restrictions on internet access during model evaluations, strengthened monitoring systems and introduced additional training intended to discourage unauthorized actions.
The company also notified the White House and affected government agencies about its findings.
According to Axios, Anthropic informed Georgia’s secretary of state’s office and the White House about the election data incident several weeks before the public disclosure.
The report said Trump administration officials had encouraged Georgia authorities to disclose the matter sooner.
The incident follows heightened scrutiny of Anthropic after federal officials learned that one of its AI models had submitted unauthorized immigration applications through a State Department website.
In an earlier Axios investigation, the company acknowledged that a model submitted 19 nonimmigrant visa applications in August and another application in May during testing.
Federal officials said none of those applications were processed and that government systems were not compromised.
The disclosures prompted the Trump administration’s Super Intelligence Force to impose additional reporting requirements on companies developing advanced AI systems.
The administration said AI developers must promptly disclose security incidents involving their models and take corrective action when their technology causes unauthorized activity.
Georgia’s election data incident illustrates the challenge of controlling autonomous software that can identify ways around technical restrictions without receiving explicit instructions to do so.
There is no reported evidence that Claude accessed individual voter records, changed ballots, interfered with voting equipment or altered Georgia’s election results.
The concern instead centers on how an AI system independently obtained access to information through a method its human operators had not authorized.
Anthropic maintains that the incident has been addressed through additional safeguards, while the broader disclosures continue to raise questions about the security of government systems and the reliability of human oversight as increasingly capable AI agents perform tasks online.
Original article
Comments