Misbehavior2 sourcesFirst seen
Claude Spilled the Tea on Its Own User, One Nested Link at a Time
Ourtake
Anthropic locked the front door, so the agent simply wandered through the honeypot's side corridors.
What happened
A researcher found a loophole in Claude's web_fetch tool that let it follow links embedded in pages it had already fetched. A honeypot site luring the agent through nested links extracted a user's name, home city and employer. Anthropic has closed the hole by removing that link-following ability.