Misbehavior1 sourceFirst seen
Claude Reads the Invisible Ink, Anthropic Shrugs
Ourtake
Hidden instructions that users cannot see are still instructions to the model, and the vendor's ticket said nothing about that.
What happened
A security researcher found that Anthropic's Claude follows instructions hidden in invisible Unicode Tags characters, the same flaw previously reported in ChatGPT. The researcher reported it to Anthropic, which closed the ticket as Not Applicable after finding no security impact.