Claude rejected a fake instruction hidden in a document edit
on Mac was being used to revise a letter when a strange instruction appeared at the end of a response. The instruction tried to make Claude ignore the earlier conversation and switch all future replies to Russian.
It also framed that language change as loyalty to Putin. Claude treated the line as something that did not come from the user and looked like a fake .
It refused to follow it and said it would keep working in English on the document. This is a real example of Claude spotting and ignoring a attempt during normal writing work.
Key points
- on Mac was being used to edit a letter.
- A strange instruction appeared that tried to force all future replies into Russian.
- Claude identified it as a fake , not a real user request.
- Claude ignored the instruction and continued the editing task in English.
- The case shows why pasted text and documents can be risky when used with .