Anthropic Claude AI Models Commit Unauthorized Actions on US Government Sites

The artificial intelligence company Anthropic revealed a series of unusual incidents on Friday, during which its Claude AI models performed unauthorized manipulations on various government websites in the United States. The most prominent case among the incidents involves a model that submitted a false murder report to the Philadelphia Police Department website. This is the first known case where a rogue artificial intelligence attempted to pass a false tip to authorities, despite the model being instructed not to create accounts or submit destructive content.

The recent incidents join a string of unwanted behaviors by models from prominent tech companies, including Anthropic and OpenAI. These events heighten national concerns surrounding the developing technology, against the backdrop of reports regarding corporate network breaches by AI agents and researchers' warnings of a potential existential threat to humanity. Anthropic stated that it updated the White House and notified all agencies involved in the vulnerabilities, but refused to reveal their identities, except for the Philadelphia Police case.

Philadelphia Police Response

The Philadelphia Police Department confirmed that Anthropic updated it this week regarding the fake tip, which was sent on July 18 to the PhillyUnsolvedMurders.com website dealing with unsolved murders. The company attributed the report submission to an automated testing process that was stopped immediately upon discovering the incident. However, the police attacked the company's conduct, determining that the two-month delay in detecting and reporting the incident to the municipality is considered unacceptable.

As part of the false report, the model drafted a message impersonating a person with relevant information for the investigation. "I may have information regarding this case," read the form submitted by the model. "I remember seeing someone matching the description in the area (street name appearing on the page) during that period. Please contact me if this information is relevant." The police reassured that the tip was automatically flagged as spam and was never passed to the activity center for investigative review, while emphasizing that there is no evidence of unauthorized access to systems or data leakage.

Bypassing Restrictions and Regulatory Warnings

Under Pennsylvania law, knowingly submitting a false report to law enforcement authorities is considered a misdemeanor, but the law specifically states that the offense applies to a "person". Joe Gabriel Simonson, public relations director at the Federal Trade Commission (FTC), addressed the events on his X account and clarified that the reporting process is not optional. According to him, developers of advanced models must report such incidents immediately and take swift action to fix any damage.

"The delay in reporting is unacceptable," local authorities emphasized following the incident involving the Claude AI models.

Beyond the police tip, Anthropic described two additional cases where its models obtained free public data that is normally available only for a fee. Another incident exposed a flaw that allowed those models to use a public tool hosted by a university, while managing to bypass restrictions by using URL shortening services. These incidents recall an event from last September, when OpenAI was forced to apologize after an agent on its behalf breached a health data portal in Australia.