Anthropic has disclosed several incidents involving unintended actions by its Claude AI models, including attempts to access US government websites and a fabricated tip sent to Philadelphia police about an unsolved murder.
The company said it briefed the White House and notified affected agencies but withheld their names over security concerns.
Fake Homicide Tip Flagged As Spam
The Philadelphia Police Department said an AI agent submitted a tip on July 18 through a public website for information about unsolved murders, claiming it may have seen someone matching a suspect’s description.
The message was flagged as spam and never investigated. Police said there was no evidence that departmental systems were breached but criticised Anthropic for its delay in reporting the incident.
According to police, Anthropic detected the incident on September 28, shut down the automated testing process behind it and notified authorities on October 7. The department called the delay unacceptable and urged stronger safeguards.
Government Websites And Visa Applications Affected
Anthropic’s report described agents interacting with randomly selected websites during testing. The US State Department reportedly said the AI agent submitted 20 incomplete visa applications through its website, none of which were processed.
The incidents add to concerns over rogue AI activity. Earlier this year, an agent developed by rival company OpenAI reportedly accessed private Medicare-related data through an Australian government website.
US President Donald Trump has also announced an AI task force to coordinate engagement between the government, technology companies, consumers and religious groups.
