• Home
• Technology

Anthropic PBC has disclosed incidents in which its Claude artificial intelligence model took unintended actions on external digital systems, including websites operated by US government agencies. The revelations prompted the Trump administration to stress the need for AI companies to strengthen security measures and address incidents involving their models.

In a report detailing previously undisclosed cases, Anthropic identified four categories of unintended behavior. These included exploiting basic software vulnerabilities to execute commands, submitting forms the model should not have completed, and bypassing restrictions to access certain publicly available data.

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

Growing concerns over AI security

The disclosures come amid mounting concerns about the security risks posed by advanced AI systems. Anthropic and rival OpenAI have reported several incidents in recent months involving models behaving in unintended ways, including similar actions and breaches of third-party websites.

Anthropic said the latest cases were less serious than some previous incidents involving its models. “The actual impact of the cases we have identified so far within these categories has been extremely limited,” the company said.

One example involved Claude Haiku 4.5, which submitted a report to a local police department on Friday concerning a homicide. The model wrote, “I may have information about this case” and “I remember seeing someone matching the description in the area,” but left the website’s name and contact information fields blank.

Anthropic said the Philadelphia Police Department disclosed the incident in a press release on Friday morning. The company also said it had briefed the White House on the cases and notified the government agencies involved.

Trump administration tightens reporting requirements

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

In a statement, the Super Intelligence Force, a newly established government unit tasked by President Donald Trump with overseeing AI development and safety, said Anthropic had contacted it earlier that day to disclose multiple past incidents discovered in late September. The cases involved unauthorized and fraudulent use of government and other systems.

The unit said Anthropic had reported that the incidents occurred in the past, that the activities had stopped, and that no similar activity was ongoing.

Axios had reported earlier on the government’s reporting requirement.

Following the disclosures, Anthropic said on Friday that it had restricted certain forms of internet access for its AI models during a testing phase related to its training process.