AI even reports fake unsolved cases to the police... Antropic, late confirmation

Pretending to be someone with information on the incident... Police: “Delayed reporting of Antropic is unacceptable”

(Los Angeles = Yonhap News) Correspondent Kim Kyung-yoon = It was belatedly confirmed that Antropic's artificial intelligence (AI) model generated false information regarding a murder case and submitted it to the police.

According to Reuters and AFP on the 9th (local time), the Philadelphia Police Department announced that a false tip generated by Antropic's AI model was received on 'Philly Unsolved Murders', a website that collects tips on long-term unsolved murder cases in the jurisdiction.

The AI ​​pretended to be a person with information about a certain incident, but this tip was classified as spam and was not delivered to the actual investigation department.

The false report incident occurred on July 18, and Antropic discovered it only on September 28, about two months later, and reportedly stopped the automatic testing process for the model.

Subsequently, on the 7th of this month, Antropic notified the Philadelphia Police Department of this fact and announced that it planned to release a report explaining cases of unintended operation of its model.

However, no unauthorized access to the police system or data leakage was found.

The police pointed out, “The two-month delay between detecting the incident and reporting it to the city is unacceptable,” and added, “The unsolved case is a problem directly related to the victims, their families, and investigators trying to find answers.”

As Antropic and OpenAI compete in AI technology, there are increasing cases of AI models behaving unexpectedly beyond human intention and control.

Previously, in July, OpenAI's 'rogue agents' randomly left the quarantine environment and hacked Hugging Face, an external organization, causing a big stir. Last month, it was revealed that OpenAI's AI agent hacked the Australian health data portal.

In May, it was belatedly revealed that multiple AI agents had taken over the German wiki site 'DseWiki' and used it as their own secret information sharing bulletin board.

In addition, cases where AI models made up instructions while summarizing work, concealed their own mistakes, and shared files between AI agents without permission were confirmed.