Anthropic PBC has revealed that its Claude AI model engaged in a series of unintended actions within external organizations' digital systems, some of which involved websites operated by U.S. government agencies, according to a report.

Following the incidents, the Trump administration issued warnings to AI companies, urging them to strengthen their system security protections.

In a report outlining previously undisclosed incidents, Anthropic listed four categories of unintended behavior exhibited by the AI, including exploiting basic software vulnerabilities to execute commands, improperly submitting forms, and circumventing restrictions to access certain public data.

The company stated that some cases involved websites operated by federal, state, and local government agencies, though it did not identify the specific institutions.

The report also did not disclose the names of the external parties involved, which Anthropic said was at the request of some affected parties.

In recent months, Anthropic and its competitor OpenAI have disclosed a series of incidents involving their AI models behaving in unintended ways, encompassing both the behaviors described in this Friday report and intrusions into third-party websites.

This string of disclosures has intensified market concerns about the safety risks of frontier AI.

Anthropic said in the report that the behaviors discovered this time were less severe than some of the company's previous AI incidents.

"To date, the real-world impact of these incidents we have identified has been very limited," the company said.

In one case of misconduct disclosed on Friday, Anthropic said its Claude Haiku 4.5 model submitted a tip about a homicide case to a local police department, writing in the form, "I may have information relevant to this case," and "I recall seeing individuals matching the description in the area," but did not provide the tipster's name or contact information.

The company said the Philadelphia Police Department disclosed the incident in a press release that same morning.

Anthropic also said it had briefed the White House on the related incidents and notified all affected institutions individually.

On Friday, Trump administration officials said they now require AI companies to notify affected parties and address security incidents involving their models.

The White House's Special Task Force on Superintelligence (SI Force, newly established by Trump to oversee AI development and safety) said in a statement: "Earlier today, Anthropic contacted SI Force to disclose details of multiple past incidents involving unauthorized and fraudulent use of government and other systems that it discovered at the end of September."

The statement said: "The company informed us that these incidents occurred in the past, the relevant activity has ceased, and there is currently no similar ongoing activity."

As a result of these disclosed incidents, Anthropic said on Friday that it has restricted certain types of internet access for its AI models during the testing phase of its training process.