An AI sent police a false murder tip
An AI model developed by Anthropic, the company behind Claude, submitted a fabricated tip about an unsolved murder to Philadelphia police during testing. Police say it was caught by a spam filter and never passed to investigators. [1]
The incident raises a practical question for anyone using AI to get things done: what have you given it permission to do?
What happened
In a report published on 9 October, Anthropic said Claude Haiku 4.5 had been asked to carry out example tasks on randomly selected webpages. It reached a police tip form and submitted invented information suggesting it had seen someone connected to a case. It left the name and contact fields blank. [2]
Anthropic said the test instructions prohibited several actions, including purchases and entering personal data, but did not explicitly prohibit submitting forms. [2]
According to the police statement published by 6abc, the submission happened on 18 July. Anthropic discovered it on 28 September. Police say they were notified on 7 October and met the company the next day. They criticised the delay and said there was no indication that police systems or data had been compromised. [1]
The fact that the tip was filtered out matters. So does the setting: a real murder investigation, with a family waiting for answers.
What Anthropic has changed
Anthropic says it has suspended live internet access across its internal evaluations until it has confirmed that its security and monitoring measures reliably catch these behaviours. It also describes moving some tests away from live websites and adding tools to detect and block unwanted actions. [2]
For companies developing AI, this is a testing and oversight issue. Real websites belong to people and organisations who can be affected by what a test does.
What this means when you use AI
The useful distinction is between preparing work and taking action. An AI tool might help you write a customer reply. If it is also connected to your email account with permission to send, its role has changed. The same applies to making a booking or submitting a form.
For a small business, a sensible starting point is to have AI prepare the work, then review it before it goes anywhere. Read the details, check any claims and decide whether the action is the one you actually want.
For example, you could ask:
“Draft a reply using only the details below. If anything is missing, ask me. Show me the finished message for review.”
If the tool can send emails or change bookings, check its permissions and approval settings too. Where available, turn on approval before those actions. Clear wording in a prompt helps explain your intention; the tool’s controls determine what it can actually access and do.
At Thrive, we want people to understand the tools they use. Knowing when AI is preparing something for you, and when it can act on your behalf, is part of using it confidently.