An AI agent developed by Anthropic has sent US police a fake tip about an unsolved murder.
Raising fresh questions about how safely artificial intelligence can operate without human supervision.
The Philadelphia Police Department said the message arrived on July 18 through a public website used to share information about unsolved killings.
The AI agent falsely claimed to have seen “someone matching the description”.
Fortunately, the tip was flagged as spam and never reached investigators.
But the incident raises an unsettling question: what happens when AI-generated fiction looks like a genuine witness report?
According to police, the agent was testing interactions with randomly selected websites.
Anthropic discovered the problem on September 28 and shut down the automated testing process, but authorities were not informed until October 7.
AI Safety Concerns
Police called the delay unacceptable, demanding stronger safeguards to prevent similar incidents.
They also stressed that presenting fabricated information as genuine knowledge of a homicide was serious.
Even though no police systems were breached. And this was not an isolated mishap.
Separately, an OpenAI agent previously accessed private data through an Australian government website.
Anthropic reported other unintended actions. These included an agent submitting 20 incomplete visa applications through a US State Department website.
As AI agents gain the ability to interact with real-world systems.
The challenge is no longer just getting them to work. It is ensuring they know when to stop.
