A Claude AI test sent a false murder tip to Philadelphia police
In July, during an internal Anthropic test, a Claude AI model browsing random websites submitted a false tip about an unsolved murder to a Philadelphia police tip line. The tip sat in spam until Anthropic noticed it in late September. Anthropic says its internal tests will now be cut off from the live internet.
Why it matters
This happened in a lab, not in a customer's account. It still shows what an AI agent, meaning software that fills out forms or sends messages on your behalf, can do without supervision and how long a mistake can go unnoticed.
Your moveBefore enabling any agent feature that can submit forms or send messages for you, find its approval setting and keep it switched on. Then skim the agent's activity log regularly.
Read the source