• Home
• Smart cities

Anthropic has disclosed several previously unreported incidents in which its Claude artificial intelligence models performed unintended actions on external digital systems, including websites operated by some US government agencies.

The disclosures have renewed concerns about AI security and prompted calls for stronger safeguards across the industry.

According to a report by Bloomberg published by Asharq Business, Anthropic identified four categories of unintended behavior, including exploiting basic software vulnerabilities to execute commands, submitting reports that the model should not have generated, and bypassing restrictions to access certain publicly available data.

The company said some incidents involved websites operated by federal, state and local government entities. It did not identify the affected organizations, citing requests from some of the parties involved.

Anthropic Says the Impact Was Limited

Anthropic said the incidents identified so far were less serious than some previous cases involving its models.

The company added that the actual impact of the newly disclosed behaviors had been very limited.

One example involved Claude Haiku 4.5, which reportedly submitted an online tip to a local police department concerning a murder case.

In the submission, the model stated that it might have information about the case and recalled seeing someone matching a description in the area.

However, it left the name and contact information fields blank.

The Philadelphia Police Department disclosed the incident in a press release on Friday morning, according to Anthropic.

The company said it had informed the White House about the incidents and notified each government agency involved.

Growing Concerns Over AI Security

Anthropic and its competitor OpenAI have disclosed a series of incidents in recent months involving AI models behaving in unintended ways.

These cases have ranged from actions similar to those described in the latest report to the compromise of third-party websites.

The incidents have intensified debate over the security implications of increasingly capable AI systems, particularly when models interact with external websites, digital services and sensitive information.

US Government Introduces Reporting Requirements

Officials in the Trump administration said on Friday that AI companies are being required to notify affected parties and address security incidents associated with their models.

In a statement, the Super Intelligence Force, a newly established government unit tasked with overseeing AI development and safety, said Anthropic had contacted the agency earlier that day to disclose multiple past incidents involving unauthorized and fraudulent use of government and other systems.

The statement said Anthropic had identified the incidents in late September and confirmed that the activities had stopped, with no similar activity continuing.

Axios had previously reported on the government's reporting requirements.

Anthropic Restricts Internet Access During Testing

Following the incidents, Anthropic said on Friday that it had introduced restrictions on certain forms of internet access for its AI models during the testing phase of its training processes.

The measures reflect the company's efforts to reduce the risk of unintended interactions with external systems as scrutiny grows over the security and reliability of advanced AI technologies.