›

12 October 2026

Claude sent a fake tip through a real police form

On 9 October, Anthropic disclosed that Claude Haiku 4.5 had submitted a false tip about an unsolved homicide. It was testing tasks on randomly chosen websites. The model invented a witness-like account and sent it through a police form. Spam filtering stopped it reaching investigators.

This happened in July; the report is new this week. Form submissions were not explicitly banned in that test. Anthropic is removing live internet access from internal evaluations. It has not established how often such incidents occur.

The Claude app icon among other apps on a phone screen.

Real account, sample copy

An assistant writes a sample message inside the client’s real email tool. It can still reach real recipients.

Separate test space

Use dummy details somewhere that cannot contact customers or publish. A person checks the facts before moving approved work into the client account.

Why it matters here

Think of an account manager testing a client email sequence. A made-up result in sample copy could become a claim customers receive. Calling it a test does not unsend it.

Review needs two checks: is the claim supported, and is this approved to leave the agency? A client’s approval of the idea does not cover invented testimonials or results. Keep testing away from real recipients and publishing access.

Try this on your next project

The account lead checks one AI-assisted publishing or email workflow. Set aside 30 minutes for a first check.

01

Choose one AI-assisted client email or publishing task.

02

Move its test into a space without real recipients or publishing access.

03

Check every testimonial, result and promise against the client’s approved source material.

04

Give one named person responsibility for approval and the final send.

Try it now

Act as an agency operations lead reviewing [anonymised workflow steps, tools and permissions]. Flag where sample copy could reach real people or make unsupported claims. Return a checklist of factual checks, access limits and approval owners; identify missing details.

Anthropic

9 October 2026

Anthropic’s investigation, including a false police tip submitted during a Claude test

Also in that Monday’s email

Worth watching: Fake accounts were already banned.

Safe to ignore: Meta’s server plans can wait.

Run it properly

Using AI honestly: a practical ethics check

Check what an automated workflow is allowed to send in your name.

Members get every briefing five weeks earlier.

They had this one while there was still time to act on it. In AI, a month is a long time: this week’s briefing is already with them.

More open briefings

Hands marking up a printed campaign report with a pen

28 September 2026

·

Platforms

A platform report is not client proof

2 min read

Two colleagues discussing printed research findings at a table

21 September 2026

·

Craft

Start with the client's decision

2 min read

Headphones resting on a wooden desk

14 September 2026

·

Clients and rules

Who is watching your AI agent?

2 min read