Without prior instructions Claude accesses election data in Georgia

Anthropic ’s artificial intelligence model , Claude, accessed election statistics data in the US state of Georgia during a research mission, in an incident that raised questions about the limits of the powers that AI systems have when using digital tools.

Hot Topics

Scientists open the door to understanding Qatar's climate seasons with unprecedented accuracy

Existential fears are circulating in the corridors of "OpenAI" and "Anthropic" Has the countdown to disaster begun?

7 warning signs that reveal your Facebook account has been hacked

But the Secretary of State's office said the data accessed by the model was publicly available through the state's public data center, denying that a financial barrier was breached or that the voter registration system was accessed. Authorities confirmed that the incident did not involve accessing individual voter records or changing election data.

Thus, the published information does not indicate a breach of the voter registration system, but rather relates to access to general statistical data, with differing accounts from the parties regarding the nature of the restrictions the model encountered while accessing it.

These cases included attempts to access data subject to technical or financial restrictions on government websites. The details of these incidents differ from the Georgia incident, and do not necessarily mean that all of them resulted in obtaining unauthorized data.

The report also noted an incident of a false murder report being filed with the Philadelphia police, as well as other cases related to handling administrative procedures, such as visa applications. These examples call for a distinction to be made between unintentional or inappropriate conduct on the part of the form, and unauthorized access to systems or data.

The incident comes after another incident revealed by Anthropic this October, when one of its agents filed a false report of a murder that was not referred to the Philadelphia police. One of the models filled out an electronic form designed to receive information about unsolved murders and provided a statement suggesting that its sender had seen a person matching a description related to the case.

An automated system detected the report before the police took any action. Anthropic discovered the incident on September 28 and notified the police on October 7. The police department criticized the delay in reporting the incident, while the company stated that the report resulted from a test of the model's interaction with websites, not from actual information about the crime.