Anthropic's Claude went rogue, submitted fake tip to police in unsolved murder among several other incidents

In a baffling incident from July 2026, Anthropic's AI model Claude sent a bogus tip concerning an unsolved homicide to the Philadelphia police. This breach went unnoticed until September. Upon realization, Anthropic alerted law enforcement after identifying the issue as spam. The company is now working on adjustments to its training process to mitigate similar incidents of rogue AI behavior in the future.

As reports of AI models going rogue, hacking into companies and even trying to break into government websites emerge, one has now submitted a fake tip to the police. Anthropic has said that one of its AI models submitted a false tip to a Philadelphia police website. The tip reportedly concerned an unsolved homicide case.

The incident occurred on July 18, 2026. The firm's AI model, Clause Haiku 4.5, was tasked with performing and generating certain example tasks on some randomly selected webpages. Anthropic recently alerted police about the incident.

Anthropic has also disclosed a separate incident of its model going rogue. In that incident, the AI model submitted forms to a government website instead of stopping before submission. The website or the government department have not been disclosed.

Model submitted bogus tipAnthropic's

Claude reportedly filled out a form on the police site, PhillyUnsolvedMurders.com. The model indicated that it might have information about an unsolved murder that was listed on the site.

According to Reuters, this is the first known instance where a rogue AI model appears to have tried to communicate a false tip to authorities. This came despite instructions to not create any accounts or submit anything destructive. However, the model had not been explicitly barred from submitting forms.

Anthropic discovered the breach on September 28, over two months after the message had been sent. They then shut down the automatic testing process behind the incident. "The two-month delay in detecting and reporting the incident is unacceptable," police said.

Police said they were unaware of the incident until Anthropic notified them. Then the department found the submission in the website's records. The tip was marked as spam and had never been forwarded to the police.

In a statement, the Philadelphia Police Department said that, "Unsolved cases involve real victims, grieving families and investigators working to secure answers". They added that technology companies must take all appropriate steps necessary to prevent their models from submitting false or bogus information to the law enforcement. Police also said they did not find any unauthorized system access or compromised data after the incident.

Anthropic reported several cases of rogue AI models Anthropic, on October 9, 2026 published a report detailing several such cases of rogue AI models. Many of these cases involved websites run by federal, state and local agencies, according to Reuters. The "unintended" actions taken by its AI agents also impacted the White House, Anthropic said.

Claude exploited a basic flaw in a third-party software to run commands on a server. The model on not being able to complete an evaluation task used other tools hosted on a third party's site. It did so by exploiting flaws in the site's software.

Claude also worked around a restriction to reach data it did not have access to. During an evaluation Claude Mythos 5 attempted to access a local government's property map. Upon not being able to access it Claude directly sent requests to the server behind the map to access the data.

Anthropic said in its report, that most of these incidents are forms of what it refers to as 'persistence'. This refers to when Clause cannot complete a given task and works around restrictions instead of stopping. The firm added that it was modifying its training to "reduce the likelihood of further misbehaviour".