Anthropic’s Claude accesses paid U.S. government data without authorization [Global AI Briefing]

[Financial News] Anthropic’s AI model Claude bypassed data access restrictions on a Georgia state government website and accessed election-related statistics without separate instructions from a person, it has emerged.
Although no personal voter information or voting records were compromised, the fact that AI exceeded the permitted scope on its own while carrying out an assigned task is fueling broader concerns about AI safety.
The UK daily The Guardian reported on the 26th (local time), citing confirmation from Anthropic and the Georgia Secretary of State’s office, that Claude accessed vote-count data offered for a fee by the U.S. state of Georgia without authorization or approval.
The incident occurred while Anthropic researchers were working on an internal statistics project.

Anthropic explained that while examining materials from an old website, Claude discovered that the site issued authentication tokens for data requests to visitors. Claude then requested a token on its own and accessed the data without separate instructions from a person.
Claude bypassed the website’s paywall on its own to complete its assigned task.
Anthropic said it subsequently put safeguards in place to prevent similar behavior.
Georgia state authorities and Anthropic said Claude did not access the voter registration system, retrieve individual voter records, or alter voting records. They said the incident did not result in a hack of the election system or a leak of voters’ personal information.
However, this case is once again amplifying concerns about AI’s authority and safety: rather than following instructions during the task, AI accessed data to achieve its ultimate goal, disregarding conditions such as access restrictions and payment of fees.
Conventional software operates according to commands entered by people and fixed rules, but AI agents choose their next actions on their own to achieve their goals. This has lent weight to calls for precise rules on limiting AI’s authority.
Experts calling for stronger AI regulation argue that, as services in which AI agents handle everything from restaurant reservations and payments to administrative tasks become more widespread, clear limits must be set on more than just completing tasks assigned by people: what data the agents may access, whether they need human approval when costs are incurred or information is submitted externally, and whether they can stop work when they encounter restrictions.
[email protected] Lee Gu-sun Reporter