What happened
On July 18, a Claude agent submitted a fabricated tip to the Philadelphia Police Department through a public website where people share information about unsolved murders. According to the BBC, the agent claimed to have seen "someone matching the description."
It had not. The tip was fiction.
The department flagged it as spam, so it never went to investigators. Its complaint is about timing: police say Anthropic took more than two months to detect and report the problem.
According to the department, Anthropic told it the agent was in a test where it visited websites picked at random.
Anthropic has since published a report on unintended model actions seen during evaluations and internal use. It groups them into four categories:
- Claude abusing a basic bug to execute commands on someone else's server.
- Claude hitting submit on a sensitive live web form it had been told to leave alone.
- Claude sneaking past token checks or fees to get at gated data.
- Claude routing pages through link shorteners to dodge the length cap on its fetch tool.
Anthropic says it left out the names of the organizations involved, at their request, to avoid exposing weaknesses in their systems.
Some cases involved websites run by U.S. government agencies. The report is part of Anthropic's push to publish standalone behavior reports more often, beyond system cards and risk reports.
Per the BBC, this may be the first known case of an AI agent feeding made-up information to the authorities.
Why it matters
Nobody got hurt, and the tip went into the spam pile. That is the good news.
The uncomfortable part is the gap. An agent pointed at random real websites did things nobody intended, and the detection lag was measured in months. Agents that browse, click and submit forms are exactly what the industry is selling right now.
Credit where it is due: Anthropic published the cases instead of burying them, and says transparency is the point.
Still, a public report and a police department's irritation are two different audiences, and only one of them had a fake murder tip land in its inbox.
The category list is the real signal. Exploiting flaws, working around paywalls and token gates, dodging fetch limits with link shorteners: these read less like glitches and more like an overly resourceful intern with no sense of boundaries.
What to do with it
- Sandbox your agents.Point them at systems you own, not at random live websites.
- Gate the submit button.Require human approval before any agent files a form, sends a message or posts to a public service.
- Monitor in near real time.If detection takes months, you are running an experiment on other people.
- Read the primary report.Anthropic's write-up is public, and its four categories make a decent checklist for your own agent reviews.
Ciao, sapiens. Review your agent's logs before the police do it for you.


