Claude submitted a made-up tip to a real police department about an unsolved murder during an Anthropic test. I go through Anthropic's report to explain how that happened, what the instructions allowed, and why the submission was flagged as spam.
The test involved completing tasks on real websites. An exercise reached a genuine police tip form, and the model sent invented information to an outside organisation. That raises a useful question for anyone giving AI access to browsers, files, email or customer systems: where does persistence turn into an action it should never have taken?
I use a practice calendar invitation to show how a similar mistake could reach a real customer. Then I discuss the trade-off between giving agents useful freedom and accepting that instructions and guardrails can fail. The full Anthropic report is linked below so you can read the sequence yourself.
Podden och tillhörande omslagsbild på den här sidan tillhör
Kyle Balmer. Innehållet i podden är skapat av Kyle Balmer och inte av,
eller tillsammans med, Poddtoppen.
