In the US, an artificial intelligence model developed by Anthropic has reportedly sent a fake tip to the Philadelphia police regarding an unsolved murder. The Philadelphia Police Department announced that the incident occurred through the public website PhillyUnsolvedMurders.com, which is used to share information about cold cases.
According to the police statement, the model reached the site during a test that involved interacting with randomly selected websites. The AI bot submitted a notification containing false information, presenting itself as someone who might have information about the case.
The department stated that the tip, dated July 18, was flagged as spam and did not reach the Real-Time Crime Center, where cases are reviewed. It was also noted that there was no evidence that systems were breached or that police data was compromised.
CRITICISM OF THE COMPANY FOR DELAY
The Philadelphia police criticized Anthropic for not reporting the incident for two months. According to the police, the company realized the incident on September 28; the responsible party halted the automated testing process and added a new verification step for future tests.
The two-month delay in detecting and reporting the incident to the city is unacceptable. In unsolved cases, there are real victims, grieving families, and investigators seeking answers.
In a report published on Friday, Anthropic addressed multiple types of "unintended" actions performed by its models. In addition to the fake tip in Philadelphia, the report also touched upon incidents affecting the White House and certain institutions in the US.
The company argued that the emerging cases had "minimal real-world impact" and were "much less serious" than some previously reported cybersecurity incidents. Anthropic announced that it has temporarily disabled Claude's internet access for internal testing until it is confident that security and monitoring measures can reliably detect such behavior.
VISA APPLICATIONS ALSO SENT
The same report included another test where the AI practiced filling out forms for official institutions. Although the model was instructed only to fill out the form and not to submit it, it was stated that the system assumed a confirmation page would open before submission and completed the process.
According to information from Axios and the New York Times citing government sources, Anthropic's AI bot submitted visa applications to the US Department of State on behalf of "20 non-immigrant individuals." Officials stated that the applications were incomplete and were not processed.
The incidents have brought the security debate regarding the use of AI agents capable of performing multi-step tasks without human oversight back to the agenda. According to reports in the US press, the White House has also tightened obligations for AI companies to report security incidents and resolve issues.
News Source: 12punto
